Specialized Security · Container Reader · Build Options

Build Routes

Given the good news — well-lit site, multiple cameras — here's how long each way to build it takes, and why they're a sequence, not a either/or choice.

The one thing that shapes everything

Build the platform once. The OCR engine is a swappable part.

~80% of the work — RTSP ingest, the in / out ledger, review queue, dashboard, TrackTik push — is identical in every route. Only the reader plugs in differently. So you can launch on a bought engine now and drop your own in later with no rework of the platform.

The three routes, timed

wk 024681012
A
Platform
Buy
4–6 wks
B
Platform
Build + tune
6–9 wks
C
Platform
Train our own
8–12 wks
Shared spine — same in every route The reader — the only part that changes

What each route means

Route A4–6 wks
Buy the readerStart here
  • buyPlate Recognizer container OCR — Docker on-prem or cloud, reads code + check digit + confidence out of the box.
  • buildThe platform: ledger, review queue, dashboard, TrackTik, multi-camera voting.
Fastest & safest. Client live soonest. Ongoing per-camera subscription. Accuracy proven by the vendor.
Route B6–9 wks
Hybrid — open core + cloud backup
  • buildOur own OCR from open-source (PaddleOCR + a container detector), tuned on real site images.
  • buyCloud OCR (~$1.50/1k) only as a fallback on low-confidence frames — pennies.
We own the IP. Near-zero licence. Resellable as our product. Accuracy leans on the multi-camera + check-digit fusion.
Route C8–12 wks
Full DIY — train our own
  • buildDetector + recognizer trained on our own collected site images. Zero vendor.
  • growsKeeps improving as we log more containers.
Total ownership. $0 licence, our compute. The product-company play — highest effort up front, best long-term.

Build once · swap the reader

Shared spine

built once · same in A, B & C
  • RTSP ingest from the Dahua cameras
  • Frame trigger + sharpest-frame pick
  • Multi-camera / multi-frame voting
  • ISO 6346 check-digit validation
  • Event store — Postgres + snapshots
  • In / out ledger + search
  • Low-confidence review queue
  • Dashboard + CSV export
  • TrackTik API push

The swappable reader

the only part that changes per route
  • A: Plate Recognizer container engine
  • B: open-source model + cloud fallback
  • C: our own trained model

Because the reader is one module behind a fixed interface, moving A → B → C later is a swap, not a rebuild.

Off-the-shelf parts to complement

LayerBuy (paid)Or free / open
Ingest + trigger Keep Dahua DSS Frigate — open NVR, does motion/object triggering free
OCR reader Plate Recognizer / Vaxtor PaddleOCR · YOLO · Hugging Face container models
Logging platform We build it — this is the differentiator and our IP. No off-the-shelf "container in/out ledger → TrackTik" exists.
Integration TrackTik REST API — straightforward push.

Recommendation

Do both — in this order

Launch on Route A (~4–6 wks), collect real site images, then swap our own engine underneath — zero platform rework.

The client is live and you're paid inside 6 weeks on a proven engine. Meanwhile every read logs a labeled image, so by the time we've got enough data, dropping in Route B / C is a module swap. You get paid fast and own the IP — you don't have to choose.