Estate-level ISR
for civilian security

0. Executive summary ↑

South African estates already spend heavily on security — and now we can prove it with a primary source: Silver Lakes Golf & Wildlife Estate (Pretoria, 1,645 erven) published its AGM budget — security opex R15.94m/yr ≈ R1.33m/month ≈ 36% of levy revenue, with the guarding contract alone (Bidvest Protea Coin) at R14.18m/yr, the single biggest expense line in the budget (research/01). ARC's member estates average ~32% of opex on security (~R320k/month) (research/03). Estates also own a lot of cameras — real fleets run 40–250+ (Kyalami ~250, Copperleaf ~150+, Izinga 116 — research/02). What they don't have is a way to make those cameras work as one system: cameras don't know each other's position, alerts are single-camera, and there is no shared map, no pattern-of-life, and no way to prove where the gaps are.

BirdsEye is a camera-agnostic software layer that sits on top of an estate's existing CCTV and produces (1) a live bird's-eye map of people and vehicles, (2) pattern-of-life over time (routes, heatmaps, dead zones), and (3) an actionable gap/weakness report — without ripping out a single camera.

The opportunity is real, but the demand story has changed shape since v0.2:

• Crime is falling steeply; fear is not. Murders fell 10.6% to 24,692 (FY2024/25) and every quarter of FY2025/26 fell further (≈23,100 est.); house robbery −20.4% in Q4. Yet Gallup 2025 ranks SA lowest of 144 countries for feeling safe (33%), and Stats SA shows self-protection behaviour rising (39.9% → 43.3%) as victimisation falls (research/15). The demand engine is the fear–data gap, not the crime rate — §2 is re-anchored accordingly.
• The cost pressure is structural and accelerating. Grade C metro guard minimum wages rose ~R5,036 → R7,003/month (2024 → Mar 2026), +35–40% in two years vs CPI ~3–4% (research/12). Manpower-heavy security gets ~7%/yr more expensive by default; software leverage gets relatively cheaper.
• ~462,967 estate properties (6.6% of SA homes, 18.1% of residential value — Lightstone deeds data, Feb 2024) in a nesting set of ~5,000–8,000 gated estates, plus ~37.6k registered community schemes (CSOS; true universe ~70k). ~70% of the market sits in Gauteng + Western Cape (research/03).
• The AI-video-analytics segment remains the fastest-growing layer (~US$34–38bn globally by 2030, ~20% CAGR), and MEA is under-penetrated whitespace.

The defensible wedge — restated honestly after the latest research. Every scaled leader (Verkada, Rhombus, Avigilon Alta, Flock) is hardware-first. The closest SA player, Vumacam, is no longer a future threat: it is already selling a "no rip-and-replace" overlay into estates ("The Network Effect", Feb 2026), runs a monitoring arm, and is Flock Safety's confirmed SA reseller (research/06). But Vumacam's product is structurally inseparable from pooling estate data into a rented, city-wide network — the one axis it cannot follow BirdsEye on. The unoccupied combination remains real: estate-owned data + camera-agnostic + map-first pattern-of-life + POPIA-native + multi-estate without the public LPR grid. Individual capability pieces exist elsewhere; that combination, productised for an HOA without a systems integrator, does not (research/11).

The honest risk (unchanged, now better evidenced): the hard part is not detection — it's cross-camera identity/tracking and map calibration, which are environment-limited. Field data hardened this pass: real-world trackers collapse to IDF1 46–62 (vs 77–80 in lab), and domain shift alone can cost 50+ mAP (research/10). BirdsEye is therefore positioned as an operator-assist tool (human-in-the-loop, within a response window), not an autonomous "who-went-where" ledger. Over-promising autonomous tracking is the single most common way products like this fail in the field.

Why this team (§19): BirdsEye is civilian ISR — and founder George Fox ran military ISR at scale (directed $455M of UAS delivering persistent ISR across Helmand). Co-founder Logan Audie built LinqProtocol, a GPU-orchestration compute network that is live with paying customers — giving BirdsEye inference at up to ~82% below hyperscaler cost. The founders also run Vtero.ai, a revenue-generating Cape Town AI consultancy (production AI systems with a 90-day guarantee, POPIA-compliant, ISO 27001) — BirdsEye is built and sold as a Vtero.ai product offering (§19, §24.1). v0.3 addition: this related-party advantage is a diligence item as much as a moat — §24 structures it arm's-length with a hyperscaler fallback case.

Funding (v0.4 — bootstrap, no external raise): BirdsEye is funded by Vtero.ai consultancy cash flow, not investors. Vtero funds the MVP build and 2–3 paid pilots; the first product revenue is paid ≤R100k trustee-resolution pilots (PMR 21(3)(a)/MOI authority, §11.3), converting to AGM-anchored annual licences. Costs, monthly burn and the peak funding need are quantified in the bootstrap cost model (model/BirdsEye-Cost-Model-v0.1.xlsx — see §20; do not hard-code a figure here). Year-1 target: 5–15 paying estates (§21). No raise is planned at any stage; an optional spin-out is considered only if the product outgrows the consultancy (§20, §26).

1. One-liner ↑

"We don't sell cameras. We make your cameras work as one system — so you can see how people and vehicles move through your estate, where your blind spots are, and close them."

2. Problem ↑

Estate control rooms already have lots of cameras and motion alerts, but:

Core insight: Cameras without geography are sensors. Sensors with geography, time, and tracks become situational awareness.

Why now — re-anchored on the evidence (v0.3 rewrite, research/15):

• Fear is decoupled from data — and fear is the buyer. SA is the lowest of 144 countries on "feels safe walking alone at night" (33%, Gallup 2025). Stats SA: self-protection behaviour rose 39.9% → 43.3% while victimisation fell. Santam: 50% of customers cite crime as their top risk while its own crime claims stabilised. Estates buy demonstrable safety — a board needs to show residents improvement, which is exactly a map + a monthly gap report.
• Manpower inflation is the budget forcing-function. Guard minimum wages +35–40% in two years (NBCPSS determinations), with structured increases to 2027. Guarding is ~⅓ of a typical estate budget; every AGM the labour line grows ~7% and the trustees look for leverage. Technology that makes existing guards faster sells into that meeting.
• Crime is still extreme in absolute terms and estate-relevant crimes persist. ~23,100 murders (FY25/26 est.) is still ~63/day; kidnapping is rising (+15.8%, and Gauteng carries 54.8% of the national total); gate-follow hijackings and insider-MO robberies hit flagship estates (Midstream, Centurion corridor, Aug 2025 reporting). GP pitch: kidnapping + gate hijackings. WC pitch: semigration-fuelled estate growth + ambient fear (gang murders +58% H1 2025 — not inside estates, but it moves the semigrating buyer).
• The private-security sector keeps growing regardless of the crime curve. PSIRA 2024/25: 637,675 active officers, 17,146 active businesses — active businesses +109% since 2015 (the durable trend; the +10%/+13% single-year deltas sit on PSIRA's restated baseline, ≈+4–5% against the prior AR series) (research/15).
• Load-shedding is over as a live shock (~26 hours of cuts in 2025) — kept here as history only: it proved estates upgrade tech in response to specific security shocks (insurers recorded ~40% burglary increases during outages). The successor shock is infrastructure sabotage/cable theft (20k+ cases/yr) — which flows to infrastructure security, an adjacent vertical, not the estate beachhead.

3. Solution — BirdsEye ↑

A software platform that:

1. Ingests existing cameras (ONVIF / RTSP / common NVRs and VMS — Milestone, Cathexis, Hikvision, Dahua).
2. Places them on a bird's-eye map of the estate (site plan / satellite / drawn map) with field-of-view wedges.
3. Detects people & vehicles and builds tracks (paths over time).
4. Aggregates into pattern-of-life: ingress, egress, corridors, dwell, weak zones.
5. Shows live + historical movement on one map for the control room.
6. Produces actionable recommendations — "blind corridor between Cam 12–14", "fence-climb cluster 02:00–04:00 west wall".

Not a camera brand. The brain + the map on top of cameras they already own.

North-star UX: the operator opens one map, not 64 tiles. A threat is an icon + trail + nearest cameras, not "motion on channel 47."

The capability ladder (v0.3 — sell in this order, per research/10):

1. Map + coverage + detection (works day one, no heroics)
2. Identity-free pattern-of-life — heatmaps, dwell, after-hours outliers (the co-hero; this is what Avigilon UMD / BriefCam / Verkada heatmaps actually sell)
3. Vehicle multi-cam + choke-point ANPR — the hard-ID plane (far more reliable than person ReID)
4. Operator-assist short-horizon continuity (30–120 s "probable same target" suggestions, human-confirmed) — a product story, not a disclaimer
5. Person cross-camera ReID — last, forensic-first, always scored, never silent

4. Product vision ↑

Core modules

Packaging (illustrative tiers)

• BirdsEye Watch — map + live icons + alerts
• BirdsEye Pattern — heatmaps, ingress/egress, weak-zone reports
• BirdsEye Command — multi-estate + incident learning loop + response-provider integrations

5. Product architecture (overview) ↑

Cameras / NVR / VMS (existing, mixed vendors)
        │  RTSP / ONVIF Profile S · Milestone AI Bridge · Hikvision/Dahua SDK
        ▼
Edge AI node(s) on site  ──►  detect (YOLO) · track · ANPR at choke points · homography → map coords
        │  tracks + events (metadata, not raw video, leaves site)
        ▼
BirdsEye core (API + DB + time-series of tracks + incident store)
        │
        ▼
Map engine (site plan + camera placement + FOV + ground-plane projection)
        │
        ▼
Apps:  Control Room  ·  Estate Manager / Board  ·  (later) Resident view

Design commitments (the why is in §11–12):

• Edge-first: detection/tracking run on-site (NVIDIA Jetson Orin or Hailo-8 boxes). Only metadata/tracks — not raw video — leave the estate. Privacy, bandwidth and cost.
• Camera-agnostic: ONVIF/RTSP as the universal MVP path; vendor SDK / VMS bridges for richer metadata where estates already run Milestone/Cathexis/Hikvision.
• On-prem option + SA-resilience: important for estates wary of cloud video, for POPIA data-minimisation, and for surviving power/connectivity interruptions (a differentiator vs US cloud VSaaS — see §12, §25).
• SA-region hosting for anything that does leave site (POPIA s72 cross-border discipline — v0.3 addition, §11.1). No estate personal information (clips, ANPR reads, ReID embeddings) runs on non-SA LinqProtocol nodes — the ~82% inference discount applies to anonymised/licensed training or SA-pinned nodes only; production inference on personal data stays in-region (residency diligence item, §24.2).

Full data model and MVP scope live in TECH-SPEC.md.

6. Market opportunity & sizing ↑

Figures researched and fact-checked across two passes (RESEARCH.md, research/03). Where methodologies disagree we cite ranges and name definitions.

6.1 The market taxonomy (v0.3 — replaces flat "8,000 estates" framing)

The addressable market is a cascade, not a single count:

Growth: estate properties 320k (2016) → 463k (Feb 2024) → 490k+ (2025); estates' share of sales volume 11% → ~16%. New-build overall is shrinking (StatsSA plans passed −20.4% H1 2025) — estates take share of a shrinking new-build pie, which suits a retrofit/BYO-camera product (most of the installed base already exists).

6.2 The spend base (context, not TAM)

• Primary-source budget anchor: Silver Lakes (1,645 homes) — security opex R15.94m/yr ≈ 36% of levy revenue; guarding contract R14.18m/yr; ≈ R808/home/month on security (research/01). ARC member average: ~R320k/month on security.
• PSIRA 2024/25: 637,675 active officers, 17,146 active businesses — a 3–4× ratio over police, one of the world's largest private security workforces.
• SA CCTV/surveillance-camera market: US$405.9m (2025) → 1.31bn (2034) @ 13.46% (IMARC) / US$761.6m (2024) → 1.46bn (2030) @ 11.7% (Grand View). AI-enabled/IP cameras fastest-growing.
• Guarding economics: a 24/7 post is ~R28k/month in Grade-C metro wages alone (4 × R7,003, Mar 2026), ~R36k+ billed with overheads; full 24/7 guarding services R31,000–63,000/mo; Gauteng 12h posts bill R15,100–17,300/mo. (The older R17.5k/R22.8k four-guard figure was a lower-grade/non-metro floor.) (research/01).

6.3 TAM / SAM / SOM (bottom-up, BirdsEye's actual layer)

Modelled on BirdsEye revenue per estate (platform fee + per-AI-camera + setup), not total security spend. Assumes ZAR ~R17/US$. ARPU inputs rebased on the new camera segmentation (§14).

Expansion TAM (later, §26): MEA video-surveillance ~US$3.7–4.6bn → ~US$5.7–9bn (2029/30); global video-analytics ~US$34–38bn by 2030 @ ~20% CAGR; adjacent verticals (golf/lifestyle estates first — the highest-security-budget subset is already the beachhead's core — then retirement villages, student housing, industrial/logistics parks).

Sizing caveat: ARPU, ops-ready share and estate counts remain assumptions until the L3 census and the paid Lightstone dataset land. The cascade above is designed so each assumption can be hardened independently.

7. Who buys it — ICP & personas ↑

Primary: gated residential estates / HOAs / bodies corporate (SA first), and the private security companies (PSCs) that run many of them. Best subset: golf/lifestyle estates — the documented highest security budgets (Silver Lakes, Val de Vie, Copperleaf, Kyalami all fit). Size: hundreds to thousands of homes; 40–250+ cameras (real distribution — §14); a control room or a contracted SOC. Secondary: industrial parks, retirement villages, student housing, campuses, logistics parks, mixed-use precincts.

Why SA first: mature estate-security culture, levy-funded budgets, acute fear–data gap, and multi-estate PSCs that concentrate buying power (Bidvest Protea Coin alone protects 68+ estates — research/07).

8. Value proposition & ROI ↑

The ROI arithmetic that lands with a board (rebased, research/01/12):

• Frame price in guard-months: a single 24/7 guarded post bills ~R25–36k/month (Mar-2026 Grade-C metro wages + overheads) and guard wages inflate ~7%/yr by determination. BirdsEye Pattern (R12,200/mo) costs ~0.35–0.5 of one post — and covers the whole estate, not one chair.
• Frame price against the security budget, not the levy: Pattern is ~2–4% of a mid-estate security budget (3.8% of the ARC-average ~R320k/mo; ~1% of a Silver Lakes-scale budget); Command (R26,750, 175 cams) is ~2.0% of Silver Lakes' security spend (~R16/home/month). Setup fees (R25–150k) sit below camera-capex items estates already approve (documented: R197k storage, R303k thermal, R1.08m biometrics, R843k pilot-drone line at Silver Lakes).
• The soft ROIs, made measurable in the pilot: false-dispatch reduction, overtime avoidance, operator time-to-locate, camera-health/outage reporting.
• Insurance: narrative, not discount. No SA insurer publishes an "AI analytics = X% off" table (verified — research/08); the realistic wedge is the monitoring-verification & evidence pack (§4) feeding broker risk surveys and master-policy renewals, with a bilateral underwriter pilot at 6–18 months. The 2G/3G alarm sunset (end-2027) is the forcing event to exploit.

9. Competitive landscape & positioning ↑

The real fight in SA is Vumacam, the PSC-owned stacks, and Cathexis — not the global VMS names. v0.3 upgrades the threat picture materially (research/06, research/11).

Positioning statement (restated carefully, v0.3): the mid-market estate-intelligence layer — easier than a full Genetec project, smarter than vendor-app tiles, more trustworthy than renting Vumacam's network. The unoccupied combination: estate-owned data × camera-agnostic × map-first pattern-of-life × POPIA-native × multi-estate without the public LPR grid — productised for an HOA. Frame as a focused product category, not "nobody has a map" (capability pieces exist; the combination doesn't).

10. Moat & defensibility ↑

These compound over time — but v0.3 adds the evidence thresholds each moat needs before it belongs in a pitch (several are conditional, not yet real):

1. Data / learning network effect (conditional). Every estate's pattern-of-life baseline sharpens with use; the incident learning-loop turns each event into a better model of "normal here." Aggregated, privacy-preserving cross-estate threat patterns improve detection for all. Moat is real only when: the data-rights policy (§24.3) exists, estates have signed training rights, and cross-estate aggregation survives a POPIA purpose-limitation review. Until then, pitch per-estate learning only.
2. Switching costs. Once the site map, per-camera calibration, integrations, and incident history live in BirdsEye, ripping it out means re-doing all of it. The case archive is sticky.
3. Integration library. The hard-won, VMS/camera-agnostic glue (ONVIF quirks + Hikvision/Dahua SDKs + Milestone AI Bridge + Cathexis) is a real barrier — and exactly why integrators want to resell rather than rebuild it.
4. Channel embedding. Woven into PSC/armed-response workflows, BirdsEye becomes part of their product; displacing it means disrupting their operations.
5. Compliance-as-moat. A genuinely POPIA-native, no-facial-recognition, operator-contracted posture is both a trust advantage and a barrier to careless entrants — and pre-positions BirdsEye for the draft Gated Access Code (§11), whose automated-decision provisions (§8.10 of the draft) point squarely at our human-in-the-loop design (the binding obligation today is POPIA s71, in force; the draft would validate the design, not create it).
6. Local trust / brand. Founder domain credibility + privacy-first stance is the opposite governance posture to Vumacam's controversy record (PSIRA charges, IR investigation, "GDPR-certified" claim — research/06).
7. Inference-cost structural advantage (conditional). Running GPU inference on the founders' LinqProtocol / LinqAI network (up to ~82% below hyperscalers) turns video-AI's worst margin risk into a cost moat. Two honest caveats (v0.3): the discount is currently unverified by third parties, and it may not transfer to an acquirer — so the financial model must also stand on a hyperscaler fallback case (§15, §24.2).

11. Regulatory & compliance (make-or-break for SA) ↑

This is where careless competitors get hurt and where BirdsEye can lead. None of it is legal advice — validate with a SA data-privacy attorney before launch (scoped question list: research/04).

11.1 POPIA (Protection of Personal Information Act, Act 4 of 2013)

• CCTV footage is personal information and analysing it is processing — fully regulated. ⚠ v0.3 correction: administrative fines are capped at R10m — POPIA has NO turnover-based fine (v0.2's "10% of turnover" was wrong; confirm the pin cite with counsel — s107 sets offence penalties, s109 the administrative-fine cap). Enforced by the Information Regulator (research/05).
• Ordinary person/vehicle detection + ANPR can rely on the legitimate-interest ground (s11) — the lighter tier; this is the Vumacam precedent (ANPR + behavioural analytics at scale, no facial recognition).
• Facial recognition / biometric matching = special personal information (s26) — prohibited by default; s27 realistically requires consent (impractical) or Regulator authorisation. Product decision: keep the core non-biometric; facial recognition is a separate, gated, off-by-default feature, if ever.
• Retention — ⚠ v0.3 correction: the draft Gated Access Code's norm for CCTV footage is 7–30 days (tighter than v0.2's "30–90 days", which applies to registers/logs). Design default retention to the 7–30-day band, configurable, documented, with secure deletion (research/05).
• Structure BirdsEye as an "operator" (processor), not the responsible party (ss20–21): the estate stays accountable; BirdsEye processes on instruction, maintains s19 safeguards, supports breach notification. Cleaner liability, cleaner sell.
• Draft POPIA Code of Conduct for Gated Access (gazette GN 7415, 30 Apr 2026; comment closed 29 May 2026; still not binding as of July 2026 — the Regulator's approved register lists only the 2022 banking codes). Three provisions matter directly (research/05):
• §8.10 automated decisions: the draft explicitly names AI "suspicious behaviour" flags and ANPR auto-open/deny as automated decisions requiring s71(3)-style safeguards — human involvement. The binding source today is POPIA s71 itself (in force); the draft Code merely signals the Regulator's direction, and our human-in-the- loop design already conforms — plus one product requirement: per-alert-type "underlying logic" explainability.
• s72 cross-border: founder access from UK/UAE and any non-SA cloud must satisfy s72 — prefer SA-region hosting and lock remote-support access controls.
• The draft's omission of biometric template-storage/key rules (flagged in filed comments) is the actual state — don't cite it as a requirement.
• RICA / audio: audio recording off by default, in product and contract (v0.3).
• Subject-access / redaction: multi-person redaction workflow is an ops cost — budget it (§25).
• Municipal layer: volatile. Johannesburg's CCTV by-law (Feb 2025) was repealed (Sep 2025) after OUTA/SAPOA/AfriForum challenge; Cape Town has its own City-property CCTV policy. Track per-municipality rules where pilots land.
• No Regulator enforcement action against estate CCTV/ANPR exists yet (verified NOT FOUND) — pressure arrives via codes/guidance, not case law. First-mover compliance is still cheap.

11.2 PSIRA (Private Security Industry Regulation Act 56 of 2001)

• "Security service" nets widely: s1(i) (monitoring signals), s1(b) (paid security advice — arguably hits the gap/weakness report), s1(h) (installing/servicing — hits edge-box work), s1(m) ("creating the impression" — a marketing-claims trap).
• The software-line is genuinely untested: case law reads the Act expansively, but no authority addresses a pure software vendor (verified NOT FOUND — research/04). The attorney opinion is therefore not optional; 15 scoped questions are ready.
• Customer-side criminality (new, v0.3): s38(3)(g) makes an estate criminally liable for carelessly contracting an unregistered provider, and s20(3) voids inconsistent contracts — registration status is a sales/diligence issue, not just a fine risk. Expect trustees' attorneys to ask.
• The in-house carve-out is the channel design spec: an estate self-monitoring with its own staff is not a security business (Code ch 4; Bertie Van Zyl CC 2009) — so "software-only + estate's own/PSC's registered control room" is the lightest lawful shape. Market evidence: Olarm's "monitoring provided by your armed-response company" model is the clean comparable.
• Amendment Act 18 of 2014 (51% SA ownership-and-control) is signed but still uncommenced — dormant, but a structuring consideration for foreign-heavy cap tables.
• Decision (unchanged in substance): ship software-only, let the estate's existing PSIRA-registered monitor/PSC run the control room; hold a registered-SOC subsidiary in reserve as a contingency.

11.3 Body corporate / HOA buying mechanics (STSMA + CSOS)

• Schemes run under STSMA (Act 8 of 2011) and CSOSA (Act 9 of 2011). Members approve the annual budget at the AGM; trustees run day-to-day within it.
• ⚡ v0.3 upgrade (from research/14): trustees can buy a pilot without an AGM. PMR 21(3)(a) special contributions for necessary, unbudgeted expenses need only a trustee resolution; PMR 21(3)(b) adds a 10% mid-year bridge. This reframes the sales cycle: ≤R100k trustee-resolution pilot mid-year → AGM-anchored annual licence (AGM season is Mar–Jul; board-endorsed budgets pass ~99%). ⚠ Scope caveat: PMR 21(3)(a) binds sectional-title bodies corporate; the golf/lifestyle beachhead (§7) is mostly HOA (NPC/constitution-governed), where no-AGM spend authority depends on the MOI's caps (verify per target) and whether a discretionary pilot is a "necessary" expense (attorney item, §28.2).
• Mounting cameras on common property is a PMR 29 "improvement" (special/unanimous resolution gates). In practice the fights are about privacy proportionality, not procedure — the case trustees read is Phillips v Bradbury [2025] ZAWCHC 430. Software overlay on existing cameras avoids the PMR 29 gate entirely — a real sales-cycle advantage over camera-led competitors (12–24+ mo cycles).
• Arm trustees with POPIA-compliance and proportionality documentation (§24.4 pack). CSOS is the dispute backstop; the body corporate is the POPIA responsible party and must register an Information Officer.

11.4 Litigation, dual-use & international

• Vumacam v JRA (2020) was decided on administrative-law (wayleave) grounds and did not validate mass surveillance — the privacy/constitutionality question is open. Design minimisation-first so the product degrades gracefully to non-biometric analytics.
• Dual-use / NCACC (new, v0.3): if the ISR heritage is marketed, keep a short legal memo on dual-use/optics risk (NCACC-adjacent) — one page, attorney-reviewed.
• EU expansion (later): avoid real-time remote biometric identification entirely (AI Act Art 5 bans it for law-enforcement use in public spaces since 2 Feb 2025; for a private operator it is Annex III high-risk and a GDPR Art 9 problem — avoid either way); post/retrospective biometric ID is high-risk with obligations phasing to Dec 2027/Aug 2028. Non-biometric detection/tracking + ANPR sits outside the Annex III biometric categories; in the EU we would run it under GDPR legitimate interest + a DPIA.

12. The honest technical reality ↑

Split the roadmap by risk, and manage expectations ruthlessly. (Updated with research/10 field evidence.)

Capability-honesty table (what we claim at MVP / at 12 months / never) lives in research/10 — sales and marketing must stay inside it.

Compute & integration economics:

• Edge hardware ladder: Jetson Orin Nano ($249/67 TOPS) → Orin NX 16GB (~$600–900/157 TOPS) → AGX Orin ($1,999/275 TOPS); Hailo-8 26 TOPS ~$220 — and Hailo-8 ReID benchmarks at 1.77 ms/crop (~1,000 FPS), so edge ReID embedding is cheap; the constraint is elsewhere (research/10).
• Realistic streams-per-box for a full detect+track+ReID+projection pipeline are far below detection-only headlines (Orin NX16 ≈ ~11 cams @15 fps per-stream vs ~40 @5 fps shared).
• Per-site compute: ~$4–8k (40 cams), ~$20–45k (200 cams) — but compute is the small line item. The dominant cost (and the moat) is integration + calibration + tuning + ML-ops.
• Ingest: ONVIF Profile S + RTSP universal MVP; richer metadata via Milestone AI Bridge / MIP SDK or Hikvision/Dahua SDKs. Do not rely on ONVIF Profile M as universal metadata (v0.3 softening).
• False-alarm budgets (new, v0.3): set per-zone-class FA targets (gate vs perimeter; thermal decision tree) and a pilot acceptance test pack — detection rate, FA rate, map RMSE, 4-hour power-loss test, VMS event visibility (§23.3).

Full MVP scope, data model and pipeline in TECH-SPEC.md.

13. Technology roadmap (12 months) ↑

MVP definition (must ship): operator sees the estate map, all cameras placed, person/vehicle alerts appear as map markers, a 7-day movement heatmap, and an exportable "weak zones" list — with retention/role controls and no raw video leaving site.

Resilience (v0.3): edge nodes must ride out power/connectivity loss gracefully (local buffering, store-and-forward, camera-health reporting) — a differentiator vs US cloud VSaaS, and the 4-hour power-loss test is part of pilot acceptance (§23.3).

14. Business model & pricing ↑

Price per estate against the guarding budget — never per camera against cloud storage. Evidence base now includes real estate budgets and estate-software price points (research/01, research/12): SA estate software transacts at R700–1,600/mo/estate (My Estate Life, Gatebox), so BirdsEye at ~3–40× the estate-software budget must be sold against the security budget (one 24/7 post = R25–36k/mo). Armed response (~R700/household/mo, flat since 2022) is the WTP ceiling per home — pitch guarding budgets, not response fees.

Revenue streams

Camera segmentation (v0.3 — replaces 20/40/200, from research/02)

Real estate fleets: Kyalami ~250 (994 homes, 8 km perimeter), Copperleaf ~150+, Kingswood ~140 planned, Izinga 116 (42 thermal / 62 optical / 12 ANPR), Val de Vie 112+ thermal. ANPR ≈ 10% of fleet. Perimeter density 6–9 thermal cams/km.

Illustrative price card (ZAR, ~R17/US$)

• Setup: R25k (small) → R150k (large), reflecting calibration/integration labour.
• Multi-year discount ladder — ⚠ v0.3 correction: evidence says annual prepay 10–20%, 2-yr 10–12%, 3-yr 15–25%; SA security's own norm is 3–6 free months (≈8–17% effective). v0.2's "20–30% industry norm" was aggressive. Use 10–15% annual / 15–20% 3-year, plus a ≥5% (CPI-linked) annual escalator (research/12).
• Only cameras with analytics enabled are billed per-camera.
• ⚠ Small-scheme warning (v0.3): at R999–1,800 levies, even Watch is 5–12% of a small scheme's security envelope — the sectional-title tier needs a PSC-bundled or stripped-down offer, not the standard card (research/01).

Every price here remains a hypothesis to validate in pilots against real budgets.

15. Financial model & unit economics ↑

Founder assumptions, not forecasts. Built to be stress-tested. Benchmarks from research/12; a full monthly cash-flow spreadsheet is a separate deliverable (§28).

15.1 Per-estate unit economics (mid, 80-cam Pattern @ ~R146k ARR)

Takeaway: the model only works through channel sales, multi-year contracts, and land-and-expand — direct one-off estate sales at these ARPUs won't clear a healthy LTV:CAC alone.

15.2 Illustrative 3-year build (⚠ reconciled)

Camera cohorts rebased to the §14 segmentation (~80 cams/estate blended) — v0.2's 500–2,000/4,000–10,000/20,000–50,000 contradicted its own 40-cam anchor. The raise narrative now matches the model: ~R27–69m ARR by Y3 (v0.2's exec-level "R50–90m" was unreconciled and is corrected).

15.3 FX & COGS scenarios (new, v0.3)

ZAR spot ~R16.5/US$ (17 Jul 2026); 2025 average ~R17.9. Model at R15 / R17 / R20 with a ±R2 sensitivity on every USD COGS line; consider FEC cover on ~50% of USD COGS once it exceeds ~R500k/yr. Note the tension: ZAR-denominated multi-year prepay increases FX risk — another reason to keep prepay discounts modest (§14).

15.4 COGS fallback case (new, v0.3)

Underwrite the model twice: (a) inference on LinqProtocol at the claimed discount; (b) inference at hyperscaler rates (the case an acquirer/diligence will run). If the business only works in case (a), the margin is a related-party artefact, not a moat — §24.2 structures the arm's-length rate card and audit rights that make case (a) defensible.

15.5 Valuation comparables (market context, v0.4)

v0.4 framing: no raise is planned — these multiples are market context for the value of the category (and any far-future spin-out conversation), not a pricing anchor. Rhombus's ~3× ARR is the realistic exit-floor multiple; Flock/Verkada are the trajectory. SA category evidence that the model monetises: AURA €13.5m Series B (May 2025), RapidDeploy acquired by Motorola Solutions (Feb 2025), iiDENTIFii $15m Series A (research/13).

16. Go-to-market ↑

Sell through the incumbents you can't out-muscle on boots-on-ground. The v0.3 GTM is channel-first, consultant-fluent, and AGM-aware (research/07, research/14).

Phase 1 — Beachhead (0–12 mo): 2–3 paid pilots (≤R100k, trustee-resolution sized — §11.3) at golf/lifestyle estates in GP/WC, or one mid-tier PSC with multiple estates. White-glove install: map + camera placement + AI on critical cams first. Sell the outcome — a monthly gap report + a control-room map. Land pilot LOIs with the POPIA/proportionality pack trustees' attorneys will ask for.

Phase 2 — Productise (12–24 mo): self-serve camera onboarding, "typical estate layout" templates, a multi-estate console for PSCs, and a partner/reseller program.

Phase 3 — Expand (24+ mo): adjacent verticals (§26), other estate-culture markets.

16.1 The channel map (new, v0.3 — ranked targets, research/07)

Watch-list (partner today, possible competitor tomorrow): Secutel (owns SecuVue), Visec/LPRCloudSA, Securitas SA, G4S/Allied. Assume Fidelity ADT is pre-empted by Vumacam (1,400 co-deployed LPR cams; Proof is its de facto control-room platform) — route around, don't chase. Additional names to verify from secondary sources: Beagle Watch (Prosegur), Stallion, SV Solutions, Pinnacle, Duxbury, Pentagon, TOPCCTV.

16.2 The gatekeeper motion (new, v0.3)

Independent security consultants (e.g. Scholtz Consulting, Rob Anderson/Adamastor, Excellerate's consulting arm) write estate risk registers, design upgrades, and adjudicate guarding tenders (Kingswood: 4 bidders incl. Fidelity/Bidvest/Thorburn, 5-yr contract). Highest-leverage motion: a free BirdsEye gap report delivered in the consultant's own risk-register format — they then spec BirdsEye into tenders. ⚠ Two compliance guards (v0.3): (1) keep the report desk/survey-based (camera positions, FOV, site plan — no footage-derived personal data) unless a short-form operator mandate + resident-notice check is in place first, else BirdsEye becomes a POPIA responsible party pre-contract; (2) the consultant or the estate's registered PSC is the adviser of record (BirdsEye supplies the software output) to stay clear of PSIRA s1(b) "security advice for benefit" — see R3, §11.2.

16.3 Channel economics (new, v0.3 — illustrative, to be negotiated)

Decide per deal: who owns install, first-line support, SLA, hardware; exclusivity vs response-agnostic multi-home (default: non-exclusive — it's the pitch to PSCs #2–#5). The "one PSC → N estates" land-and-expand math is what carries §15. ⚠ Guardrail: at the churn-adjusted LTV, the PSC motion clears 3:1 only if channel share stays ≤~30% or channel CAC ≤~R40k (at 40% share + R60k CAC it falls to ~2.0×) — cap the negotiable share, and state whether rev-share sits in COGS or S&M.

16.4 Ecosystem integrations (new, v0.3)

• Estate apps / board reporting: My Estate Life, EstateMate, WeconnectU, Glovent (130+ estates) — board-report integrations, not competition.
• Access-control event bus: Impro / Suprema / ATG Digital — gate events enrich tracks.
• Dispatch layer: Immix / Bold Manitou / Patriot / AURA (on-demand response network) — the response-provider-agnostic promise, made technical.
• Insurance/broker channel: Addsure-type community-scheme brokers via the evidence pack (§8), not premium discounts.

16.5 Sales calendar (new, v0.3)

AGM season is Mar–Jul (body-corporate FYE defaults to Feb + PMR 17's 4-month rule). Motion: mid-year trustee-resolution pilot (≤R100k) → results in the AGM pack 1–2 months pre-AGM → annual licence at AGM. Camera-led competitors need 12–24+ mo (PMR 29 resolutions, capex phasing); the software overlay runs 1–3 mo pilot + 3–9 mo licence.

17. Risks & mitigations (register) ↑

18. SWOT ↑

Strengths — camera-agnostic (no capex barrier); map-first UX; POPIA-native governance; founder ISTAR/security domain credibility; local & privacy-first vs Vumacam; channel-friendly (response-provider- and VMS-agnostic); LinqAI bench + inference-cost advantage (arm's-length structured).

Weaknesses — the hard tech (cross-camera ReID) is genuinely hard (field data confirms); no track record yet; integration-heavy (services drag on margin/cash); long, resolution-gated sales cycles; brand name legally compromised (§27).

(v0.3 fix: removed the stale "single-founder key-person risk" line — the team is two founders; key-person exposure is now correctly carried only as R9.)

Opportunities — Vumacam controversy + its pooled-data model open an ethical lane; fear–data gap keeps demand structural while crime falls; guard-wage inflation (+35–40%/2yr) strengthens the ROI story every AGM; MEA under-penetrated; AI-analytics fastest-growing segment; proven SA category appetite for AI/security builds and exits (AURA €13.5m Series B, RapidDeploy→Motorola exit, iiDENTIFii, Kalon→visionAI) validates the category without needing its capital (§20); PSC consolidation concentrates buyers; 2G/3G alarm sunset (end-2027) forces estate tech refresh decisions.

Threats — Vumacam's estate overlay + channel pre-emption (live); PSC #1 builds or locks in a rival stack; Canon/Milestone productise pattern-of-life; regulatory tightening on biometrics; ZAR/USD COGS squeeze; surveillance backlash tarring the category; DeepAlert/iSentry moving up-stack.

19. Team & why us (founder-market fit) ↑

A team that has just taken a hard, distributed-systems AI product from zero to shipped — with paying customers — led by a founder whose military career is the BirdsEye thesis.

Founders

George Fox — Founder & CEO (operations, GTM, domain). Operations executive and serial co-founder; 14+ years owning P&L, product and people across defence, offshore energy, Web3 and AI. Track record: grew revenue 32× (£250K → £8.2M) at Severn Offshore; drove valuation $200K → $60M at ARC Solutions; and, as a British Army Captain / Operations Officer, directed the deployment of $455M in UK/US Unmanned Aerial Systems delivering persistent ISR across two provinces of Helmand, running SIGINT and multi-source intelligence analysis (44 HVTs neutralised in 6 months; 37 threat vectors intercepted leading Salama Fikira's ops/intel cell). MBA (Warwick), BScEcon Int'l Politics & Military History (Aberystwyth), CAA PPL, PRINCE2, Six Sigma Green Belt, NEBOSH, Operator CPC. → BirdsEye is civilian ISR. George has run military ISR at scale.

Logan Audie — Co-founder & CTO / Product (build). Founder & CEO of LinqAI; shipped a suite of AI products — Synthetiq (#4 Product Hunt Product of the Day), Marketr, Creatr, Composr, Researchr, Designr, LinqHub — and architected LinqProtocol, a global GPU-orchestration marketplace (Kubernetes, cross-domain auth, distributed compute). → The CV/ML + full-stack + product engine, already proven at ship quality.

The proof point

LinqProtocol — a decentralized GPU compute network on Arbitrum L2 delivering up to ~82% cost savings vs AWS/GCP/Azure — is complete, live, and has paying customers. Taking a product that hard from concept to revenue de-risks the investor's biggest pre-seed question. Diligence caveat (v0.3): the ~82% figure is the founders' own claim — independently unverified to date; §24.2 structures it arm's-length, and §15.4 models the hyperscaler fallback.

The vehicle

BirdsEye is built and sold by Vtero.ai, the founders' Cape Town AI consultancy — production AI systems delivered under a 90-day guarantee, POPIA-compliant, ISO 27001, Western Cape-based (the beachhead's home turf, §6) and hosted on LinqAI. BirdsEye is Vtero's estate-security product line: contracts, IP and hiring all run through Vtero (§24.1), and what an HOA board diligences is the consultancy's trading record — not a startup deck (§20).

Unfair advantages (from the LinqAI platform)

1. Inference at cost — video-AI's margin-killer becomes a structural margin advantage (see §10.7 for the two caveats).
2. In-house AI/CV capability — a working AI-engineering org, not a hire-plan on a slide. But (v0.3): "the LinqAI bench covers it" is not free capacity — secondees carry real cost and need IP-assignment hygiene (§24.1).
3. Multi-entity, multi-jurisdiction operating experience (UK/UAE/Bahamas) and GDPR + ISO 9001/14001/31000 frameworks already run by the founders — transferable to POPIA-native operations.

Hiring plan (lean) & org by stage (v0.3)

Org shape by stage: pilot (2 founders + 1 contract ML + 1 contract SA ops) → 20 estates (+1 full-stack eng, +1 ops/install, fractional compliance) → 100 estates (+channel lead, +support FTE per §25 ratio, +ML-ops). ZAR fully-loaded costs are modelled in model/BirdsEye-Cost-Model-v0.1.xlsx (§28.6). Founder time commitment (state openly with boards and partners): each founder's % FTE to BirdsEye and the plan for Logan's concurrent LinqAI CEO role (transition / deputy / defined cap) — the same diligence question, now answered to estates and channel partners rather than investors (R9, §24.2). Advisory seats to fill early (v0.3): SA privacy counsel, a PSC operations executive, a CV/ML advisor.

20. Bootstrap funding model (v0.4 rewrite) ↑

The decision: no external raise — not now, not staged for later. BirdsEye is built and sold as a product offering of Vtero.ai, funded out of consultancy cash flow plus its own early revenue. The v0.3 staged raise (R8–12m pre-seed SAFE/CLN → R25–45m seed), investor-target list, pre-money anchors and Series-A gates are removed. Seed-round optionality is preserved (bottom of this section) but is not the plan.

Sources of funds (in order of reliance):

The cost model: all burn, runway and cohort economics live in model/BirdsEye-Cost-Model-v0.1.xlsx (workbook in progress at v0.4 drafting). Tabs: assumptions · Vtero cash-flow allocation · MVP build cost · pilot unit economics · licence ramp & cohorts · monthly burn / runway · peak funding need · scenarios (LinqAI vs hyperscaler inference per §15.3/§15.4). Peak funding need: see the model — do not hard-code a figure here. Vtero's draw on consultancy cash flow is bounded by that peak-need number and reviewed against it monthly.

Stage gates & kill criteria (founder-set; reviewed at each gate):

Why bootstrap fits this business:

• The channel sets the speed, not capital. Sales move through PSCs, security consultants and trustee resolutions (§16) — reference sites and trust close deals, not headcount blitzes. Extra cash would not compress that motion.
• AGM calendars are the revenue clock. Budgets pass at AGMs (Mar–Jul season, §11.3); a raise in the bank doesn't move an AGM date. The ≤R100k trustee-resolution pilot exists precisely to sell inside that cycle.
• No dilution, no clock. Founders keep 100% of the product upside (§15's illustrative R27–69m ARR by Year 3) and answer to trustee boards, not investor boards — no forced growth ahead of the evidence base (§23).
• Vtero.ai is the credibility. An HOA board buys from a POPIA-compliant, ISO 27001 Cape Town consultancy that ships production AI systems under a 90-day guarantee — a trust bar a standalone startup with a term sheet would still have to earn.
• The cost structure is already disciplined. Inference at cost via LinqAI (§15.3, arm's-length per §24.2) plus the existing consultancy bench keeps burn low enough for cash flow to carry it — quantified in the cost model.

Optionality (one paragraph, then the subject is dropped): bootstrapping does not burn the boats. If the product outgrows the consultancy — channel pull faster than cash flow can fund, or an international expansion case — a seed round or spin-out can be raised later from a position of revenue and pilot metrics, on materially better terms than the v0.3 plan assumed. Until then, no raise is planned and the plan carries no investor machinery.

21. Success metrics (Year 1) ↑

22. The script (elevator → board) ↑

30 seconds: "Estates have dozens of cameras that don't know where each other are. Operators get motion pings, not a picture of the crime. BirdsEye puts every camera on a map, shows how people and vehicles actually move through the estate, and shows weak points over time — so security gets a bird's-eye view and the estate gets smarter after every incident. No new cameras, and you own your data."

2 minutes (⚠ v0.3 — claims softened to match product truth): "Today the hub is reactive: motion → radio → maybe a check. BirdsEye is proactive geography: ingress, egress, busy paths, dead zones. Install a camera because the data says so, not because someone feels exposed. When something happens, the operator sees the track on the map with the nearest cameras — and the system suggests which nearby camera probably has the same target, with a confidence the operator confirms. Same cameras, better awareness. And unlike renting a city-wide surveillance network, the estate owns its own data, there's no facial recognition, and it's built POPIA-first for the board and the residents."

(Removed from v0.2: "Follow a track across the estate, not channel-by-channel" — that claim outruns the cross-camera evidence, §12. It returns when pilot metrics earn it.)

23. Demand evidence & pilot program (new, v0.3 — was the plan's biggest hole) ↑

23.1 What we already know (desk + primary sources)

• Security is the largest levy line item — proven with a primary budget. Silver Lakes AGM pack (approved Apr 2025): security R15.94m/yr ≈ 36% of levies; guarding contract R14.18m/yr; and a budgeted R843k "Security – pilot project drone" line — estates already carry pilot-sized opex lines (research/01, research/14).
• ARC member estates average ~32% of opex on security (~R320k/mo).
• Estates approve six-figure security capex at AGMs (documented R197k–R1.08m items), and CSOS precedent (CSOS3934/KZN/21) shows trustees personally liable for unapproved capex — subscription opex is structurally easier to approve than capex.
• Named estate levies/landmarks: Waterfall Country Village R3,610 (784 homes), Steyn City R6,540, Val de Vie R4,770, Midstream ~R2,121 (~5,000 families).

23.2 The discovery program (to run, pre-pilot)

23.3 Pilot design (new)

• Selection rubric: VMS type (Milestone/Cathexis/Hikvision mix), camera count 40–120, PSC relationship, trustee champion, site-plan quality, POPIA maturity, GP/WC location.
• Policy: paid pilots only (≤R100k, trustee-resolution sized — §11.3); free only for a named reference flagship with signed case-study rights.
• 14-day learn-in, then a 30–60-day measurement window.
• Acceptance test pack (pass/fail, agreed up front): detection rate on walked routes, false-alarm rate by zone class, map projection RMSE, operator time-to-locate vs tile wall, 4-hour power-loss resilience test, VMS event visibility, retention/audit working.
• Conversion path: pilot → annual licence at the next AGM, with reference rights and a board presentation slot baked into the LOI.
• Funnel targets: 10 qualified → 5 LOIs → 2–3 paid pilots → ≥50% convert.

24. Commercial, legal & data structure (new, v0.3) ↑

24.1 Entity, contracting & IP hygiene (v0.4 rewrite)

• Contracting entity: Vtero.ai (Pty — confirm suffix with founders, TBC) signs every LOI, pilot agreement and MSA. BirdsEye is a product line and trademark of Vtero, not a separate company; the v0.3 holdco/cap-table/option-pool framing is removed with the raise.
• IP: all BirdsEye product IP sits in Vtero.ai — including written assignment for any LinqAI engineers seconded to BirdsEye work. What Vtero owns must be unambiguous.
• Related-party hosting: LinqAI ↔ Vtero hosting/inference stays arm's-length per §24.2 (unchanged).
• Internal P&L and brand hygiene: BirdsEye runs as a separate product code with its own books inside Vtero — revenue, pilot costs and support load tracked against the §20 stage gates, so the bootstrap decision is made on real product economics. The trademark filing under the new name (§27) is made by Vtero.ai.

24.2 Related-party GPU (LinqAI/LinqProtocol)

The cost moat is also a diligence landmine without structure:

• Arm's-length rate card between BirdsEye and LinqAI/LinqProtocol, documented and benchmarked to hyperscaler list prices.
• Audit rights for investors on the related-party pricing.
• Hyperscaler fallback: contracts and the financial model (§15.4) both assume inference can migrate to AWS/GCP/Azure at market rates without breaking unit economics.
• Transferability disclosure: the discount may not survive an acquisition — said before diligence finds it.
• Governance of the dual role: Logan (CEO of LinqAI and BirdsEye CTO) sits on both sides — related-party contracts are approved by disinterested director(s)/investor consent with Logan recused; "arm's-length" means independently approved, not merely benchmarked.
• Data residency: the rate card names which workloads may run on LinqProtocol — no estate personal information on non-SA nodes (§5, §11.1 s72).

24.3 Data strategy & model training rights (resolves the §10.1 tension)

24.4 Commercial liability & contract stack

• LOI → pilot agreement → MSA ladder, with the trustee approval pack (POPIA operator agreement + PIIA/proportionality template + resident notice — drafts ready in research/05, attorney review pending).
• Liability: capped (fees-paid style caps); no crime-prevention warranty — BirdsEye is a decision-support tool; human-in-the-loop language in contract matching the product truth (§12).
• Insurance stack: tech E&O + cyber, sized for estate security work.
• Reference rights: case-study consent negotiated in the pilot LOI, not after.

25. Implementation, support & SLA (new, v0.3) ↑

Edge + calibration dominate cost; capacity and uptime are the operational truth behind the SaaS story.

26. Adjacent verticals & product trajectory (exit framing removed, v0.4) ↑

Vertical sequencing (after beachhead proof):

Product trajectory (v0.4 — replaces the exit/strategic-buyer map): BirdsEye is a product line of Vtero.ai, and the default path is that it stays one — growing through the verticals above as consultancy-funded product revenue. An optional spin-out or acquisition is considered only at scale, if the product outgrows what a consultancy P&L can fund (§20). The natural strategic owners of this layer are unchanged — Fidelity Services Group, Bidvest Protea Coin, Beagle/Prosegur, the Maziv/Remgro stack (conflicted), global VSaaS entrants wanting an African estate footprint — noted as market context, not as a target. M&A precedent: Motorola Solutions acquired RapidDeploy (Feb 2025) and Blue Eye (2024) — public-safety/remote-monitoring tuck-ins are active, so a strategic exit exists if the founders ever want one.

27. Naming, trademark & trust brand (new, v0.3 — elevated to blocker) ↑

Verdict from research/09: "BirdsEye" is HIGH RISK — rename before any brand spend.

• Same-industry collision: Birdseye Security Solutions (Mississauga, Canada) — AI video-surveillance/remote-monitoring, ~15 years operating, VC-backed. Near-identical category.
• Generic in the exact product category: "bird's-eye view" is the dictionary term for the core feature (Bing Maps "Bird's Eye", Frigate's literal birdseye: config key, automotive "bird view" systems) — weak registrability, near-unenforceable.
• Crowded field: Nomad Foods' Birds Eye (enforces globally), Garmin BirdsEye®, Birdeye.com (large SaaS), birdseye.com/.co.za/.ai/.io all registered.
• Status: CIPC register not publicly indexed — a special search (~R2,500) + attorney is required; SA is not a Madrid member.
• Ranked fallbacks (screened, collisions documented in research/09): 1. Eyrie · 2. Shikra · 3. Buteo (then Goshawk, Uitsig, Aquila with reservations). Killed with evidence: Kestrel, Bateleur, Horus, Vigilo, Panoptes, Augur.
• Trust-brand system (carry over whatever name wins): resident notice templates, transparency page, crisis-PR playbook for surveillance backlash (R8), ESG-consistent posture (no silent face enrolment, bias monitoring per §12/R11).

28. Immediate next steps (v0.4 — with explainers) ↑

Framing (v0.4): BirdsEye is now a product offering of Vtero.ai, bootstrapped — no external raise, ever planned. Vtero consultancy cash flow funds the build; paid pilots are the first revenue; licences and channel deals carry it from there. The v0.3 pre-seed step is deleted (§20 kept as dormant reference). Every Rand below is Vtero's own money: no spend without a deliverable, nothing outside the cost model (model/BirdsEye-Cost-Model-v0.1.xlsx, §28.6). Costs are founder estimates (ranges, MED-LOW confidence) — they seed the workbook, then true-up monthly against actuals.

Sequencing: rename + legal run in parallel with the thin-slice build (they don't block code); LOIs once a demo exists; paid pilots only after the slice passes its acceptance pack; conversions land in the Mar–Jul AGM season (M8–M12, §16.5).

28.1 Rename + trademark clearance — M1–M2

• What: pick and clear the replacement for the HIGH-RISK "BirdsEye" name (§27).
• Why: same-industry collision + generic term (research/09) — brand spend before clearance is spend at risk; pilots must market the final name.
• How:
• CIPC special search (~R2,500, §27) on the ranked fallbacks: Eyrie / Shikra / Buteo.
• Attorney clearance opinion on the winner; file the mark; register domains day one.
• Carry the trust-brand system (notices, transparency page, crisis-PR playbook) over (§27).
• Deliverable: cleared name, filed application, brand-swap checklist executed.
• Est. cost: R15–40k (search cost known; clearance + filing est, LOW-MED confidence).
• When: M1–M2 — decided before any pilot marketing (§27).

28.2 SA attorney engagement — M1–M3

• What: the fixed-fee legal package that de-risks the regulatory posture (§11).
• Why: PSIRA's software-line is genuinely untested (§11.2) and trustees' attorneys will ask for the POPIA/PMR paperwork — the opinions are a sales asset, not overhead.
• How:
• PSIRA software-line opinion on the 15 scoped questions (research/04; fixed-fee, est 2–4 weeks) — confirms the "software-only + estate's registered control room" shape.
• POPIA review of the research/05 drafts (operator agreement, PIIA, resident notice → §28.8).
• PMR 21(3)(a)/PMR 29 confirmations incl. the HOA caveat: PMR 21(3)(a) binds sectional-title bodies corporate; HOAs need per-target MOI cap checks (§11.3).
• IP hygiene for the Vtero structure: written assignment for any seconded LinqAI engineers (§24.1).
• Deliverable: written PSIRA opinion + marked-up POPIA pack + PMR memo in the data room.
• Est. cost: R35–50k (est).
• When: M1–M3 — PSIRA opinion in hand before LOIs go out (§28.4).

28.3 Thin vertical slice + acceptance pack — M1–M4

• What: build the MVP exactly as specced in TECH-SPEC.md §6, with the §23.3 acceptance pack written as executable tests.
• Why: the slice is what pilots buy; scope creep is the bootstrap killer — the NOT list (no cross-camera ReID, no facial recognition, no auto-calibration, no multi-estate console) is a funding decision, not modesty.
• How:
• Scope-lock the §6 checklist: site-plan map, 4–8 RTSP feeds, person/vehicle detect, tracks as paths, 7-day heatmap, weak-zones PDF, POPIA baseline (retention/roles/audit).
• Lab rig (dev edge appliance + test cameras), weekly demo cadence.
• Encode the §23.3 pack: walked-route detection rate, FA rate by zone class, map RMSE, time-to-locate vs tile wall, 4-hour power-loss test, VMS event visibility, retention/audit.
• Deliverable: passing acceptance run on 8 cameras + a demo reel for LOI meetings.
• Est. cost: R60–150k cash (dev edge appliance, test kit, lab install — est); people time carried by the Vtero bench, costed internally in the workbook.
• When: M1–M4 (maps §13 Q1–Q2).

28.4 Pilot LOIs — M3–M6

• What: turn a 10-estate golf/lifestyle target list (GP/WC) into 5 signed LOIs (§16 Phase 1; funnel 10 qualified → 5 LOIs → 2–3 paid, §23.3).
• Why: LOIs with the conversion path baked in are the cheapest demand proof there is — and the funnel math is what §28.5 spends against.
• How:
• Build the list against the §23.3 rubric: VMS mix, 40–120 cameras, PSC relationship, trustee champion, site-plan quality, POPIA maturity, GP/WC.
• Run the §16.2 gatekeeper motion — free gap report in the consultant's risk-register format, both compliance guards: desk/survey-based only (no footage-derived personal data pre-mandate), consultant/registered PSC stays adviser of record (PSIRA s1(b)).
• Log objections in the §23.2 log; price only as % of security labour (§14).
• Deliverable: 5 signed LOIs with reference rights, AGM conversion path and a board slot baked in (§23.3).
• Est. cost: R20–50k travel/subsistence, GP + WC (est).
• When: M3–M6 — demo from §28.3 in hand first.

28.5 Paid pilots (2–3) — M5–M10

• What: 2–3 paid pilots at ≤R100k each, trustee-resolution sized (§11.3, §23.3).
• Why: paid pilots are the first revenue and the whole bootstrap proof — free pilots teach nothing about willingness to pay; free only for a named flagship with signed case-study rights (§23.3 policy).
• How:
• White-glove install per §16 Phase 1: map + placement + AI on critical cameras first; verify each target's no-AGM spend authority (PMR 21(3)(a) body corporate, MOI caps for HOAs — §28.2 memo).
• 14-day learn-in, then a 30–60-day measurement window against the signed acceptance pack.
• Deliver the monthly gap report + control-room map; queue conversions for the next AGM, results in the board pack 1–2 months pre-AGM (§16.5).
• Deliverable: 2–3 signed pilot agreements, passed acceptance packs, first gap reports delivered, conversion decisions scheduled.
• Est. cost: R150–400k gross for 2–3 installs (edge appliance ~R30–60k each, travel, calibration labour — est); offset by R150–300k pilot fees — first revenue.
• When: M5–M10 — only after §28.3's acceptance pack passes.

28.6 Cash-flow discipline — M1, then monthly

• What: run the business off model/BirdsEye-Cost-Model-v0.1.xlsx — the single source of truth for burn, runway and pilot economics.
• Why: bootstrapped means no raise papers over sloppy cash management; Vtero consultancy revenue + pilot fees + early licences must visibly cover the peak.
• How:
• Load this section's estimates as the v0.1 budget on day one; reconcile actuals monthly.
• Monthly founder review: burn, runway, pilot unit economics, calibration hours/site (§15).
• Throttle rule: any line trending >20% over budget needs founder sign-off before more spend.
• Deliverable: workbook live from M1 + twelve monthly review notes; a current peak-funding-need figure at all times.
• Est. cost: R0–10k (bookkeeping/banking tooling, est) + founder time.
• When: M1, monthly thereafter.

28.7 Vumacam mystery-shop + Lightstone dataset — M2–M5

• What: two purchases that harden the market numbers competitors can't check.
• Why: Vumacam's estate pricing/ingestion terms are the benchmark we position against (§9), and the L2–L3 gap — how many estates are actually ops-ready — is the sizing assumption everything hangs off (§6.1).
• How:
• Human outreach (estate-manager persona) for Vumacam's current estate rate card and ingestion terms, per the research/06 gap list.
• Buy the Lightstone primary estate count/geo/value cut; rebuild the §6.1 L2–L3 layers from deeds data, not proxies; re-check the §14 price card against real software spend.
• Deliverable: Vumacam pricing memo + a hardened L2–L3 table spliced back into §6.1.
• Est. cost: R15–60k (dataset est R15–50k, LOW confidence; mystery-shop is time + travel).
• When: M2–M5.

28.8 POPIA operator pack + trustee approval pack — M2–M4

• What: finalise the attorney-reviewed operator agreement, PIIA/proportionality template and resident notice (research/05 → §24.4); print the trustee approval pack.
• Why: this pack is what trustees' attorneys ask for before a pilot signature (§11.3, §16 Phase 1) — having it printed shortens every sales cycle; it is the compliance-first positioning made physical.
• How:
• Incorporate §28.2's attorney mark-ups into the three research/05 drafts.
• Assemble the trustee pack: operator agreement, PIIA template, resident notice, Information Officer registration note, draft Gated Access Code posture (§11.1).
• Print/binder a small run for LOI and board meetings; masters in the data room.
• Deliverable: signed-off pack in the data room + printed trustee packs in hand.
• Est. cost: R15–40k (final attorney pass + assembly, est).
• When: M2–M4 — before the first LOI meetings land.

28.9 Channel conversations (top-3) — M4–M8

• What: open structured talks with the §16.1 top-3 — Bidvest Protea Coin, 24/7 Security, Verifier — on the "Switzerland layer" pitch.
• Why: mid-tier PSCs are assembling their own stacks to fight Fidelity's (§9); a response-agnostic, VMS-agnostic, non-exclusive overlay is how they do it without building — and one PSC is a multi-estate channel, not one sale (§16.3).
• How:
• Pitch: their control room, their clients, our map — non-exclusive multi-home default (§16.3).
• Bring the demo reel (§28.3) and one pilot's early results if timing allows.
• Hold the §16.3 guardrail in any term sheet: channel share ≤~30% or channel CAC ≤~R40k.
• Deliverable: three first meetings logged + one channel pilot term sheet drafted — or a written reason each target passed.
• Est. cost: R10–30k travel (est).
• When: M4–M8 — parallel with direct pilots; channels can also source them.

Definition of success at M12 (Jul 2027): 2–3 estates live as named references — acceptance packs passed, case-study rights signed, monthly gap reports in their board packs; ≥50% pilot→paid conversion into AGM-anchored annual licences (§21), with written reasons for any that didn't; the renamed brand cleared (special search + attorney opinion, mark filed); and funding that came entirely from Vtero.ai consultancy cash flow plus pilot and early-licence revenue — no external capital taken. Two numbers tell that story: this action plan's direct cash costs (R320–830k gross, offset R150–300k by pilot fees, labour carried by the Vtero bench), and the fully-loaded peak funding need in the workbook (model/BirdsEye-Cost-Model-v0.1.xlsx Summary tab — ~R2.3m base / ~R4.0m conservative with labour costed at market rates), which Vtero's cash flow must visibly cover month by month, the workbook's actuals there to prove it. A top-3 PSC in an active multi-estate pilot discussion is the kicker, not the requirement.

29. Appendix — assumptions & source posture ↑

• Verified across two passes (RESEARCH.md + research/): estate counts & Lightstone deeds data, CSOS registered schemes, PSIRA workforce (2024/25), crime stats through FY25/26 quarters, CCTV & video-analytics market sizes, competitor funding/ valuations & Vumacam scale/ownership, real estate camera fleets, one primary HOA security budget (Silver Lakes), guard-wage determinations, VSaaS & SA price anchors, POPIA/PSIRA/CSOS/EU-AI-Act framework incl. draft Gated Access Code detail, edge-hardware specs & ReID field benchmarks.
• ⚠ Corrected in v0.3: POPIA fine cap (R10m, no turnover basis); CCTV retention norm (7–30 days, draft Code); Vodacom's Maziv stake (30%, not ~35%); camera segmentation (rebased to real fleets); multi-year discount norms (10–20%, not 20–30%); NRR anchor (95–103% band, not 82%); exec-ask ARR (reconciled to §15.2); camera cohorts (rebased); load-shedding (historical only); "CAOSA" (doesn't exist — ARC/NAMA); SWOT single-founder line (stale, removed); sales-script cross-camera claim (softened).
• Changed in v0.4: funding strategy (raise → bootstrap), vehicle (standalone → Vtero.ai product offering), exit framing removed.
• Founder assumptions, still unvalidated: ARPU/price card, ops-ready (L3) estate share, 3-year build, CAC/LTV specifics, TAM/SAM/SOM ARPU inputs, channel rev-share levels, support ratios, adjacent-vertical segment sizes (retirement/PBSA flagged unvalidated in §26), the LinqAI ~82% inference discount (third-party unverified), and the bootstrap cost-model figures in model/BirdsEye-Cost-Model-v0.1.xlsx (burn, pilot economics, peak funding need — workbook in progress at v0.4 drafting).
• Open items needing humans, not web research (from research/README.md): SA attorney opinions; paid Lightstone dataset; 3–5 more HOA budget packs; Vumacam estate rate card (mystery-shop); CIPC trademark special search; estate security-manager interviews; an adversarial verification pass on the new load-bearing figures before investor use.

See RESEARCH.md for the pass-1 evidence base, research/ for the 15 pass-2 reports (start at research/README.md), and TECH-SPEC.md for the MVP architecture.

30. Appendix B — Resources & references ↑

The working library behind this plan. Every load-bearing claim in §§1–29 traces to one of these, or to RESEARCH.md and the research/ reports (index: research/README.md). Compiled 19 July 2026. Two honesty rules applied: (1) sources the research passes recorded as NOT FOUND, paywalled or dead — e.g. Lightstone article 15958 (HTTP 500), the Regulator's unpublished 2023/24 CCTV-compliance study, PSIRA guidance on software/AI, WeconnectU's rate card — are omitted, not guessed; (2) a few canonical pages not URL-cited in research/ (act texts, EDPB guidelines, vendor pricing pages) were fetched and verified live on 19 Jul 2026 — marked †. All other URLs are cited in the repo's research files.

B.1 POPIA & privacy

B.2 PSIRA

B.3 Sectional title & estates

B.4 Market & budgets

B.5 Technical benchmarks

B.6 Non-dilutive funding (bootstrap-relevant)

B.7 Competitor references

Deliberately excluded (recorded NOT FOUND / unreachable in research, 19 Jul 2026): Lightstone corporate article 15958 (HTTP 500 — the GP/WC stock-split basis, flagged as a paid-dataset gap in research/README.md); the Information Regulator's unpublished 2023/24 CCTV-compliance study; the Regulator's dedicated PIIA guidance note (due 2026/27); PSIRA circulars on software/AI scope (none exist); WeconnectU's rate card (sales-led); Fidelity/ADT quantified incident series; peer-reviewed OSNet-on-Orin throughput; official SEDFA site (unreachable — guide linked instead).