3fts

This is what it does

Try it

Visual Inventory Pro turns a shelf photo into a dated audit record. Open a shelf and compare two scans — the count, the confidence, the damage flags and the utilization history are all there.

Runs in your browser — nothing to install

  • What it does: every scan stores a box count, a 0–1 confidence and the photo it came from, on a timeline that belongs to one shelf — so a change is a diff between two photos rather than a number somebody disputes.
  • What to try: open the audit trail on one shelf and step between its scans, then look at the discrepancy and damage queues and the utilization heatmap.
  • What it runs on: the product’s own static dashboard over its own deterministic seeded warehouse — 40 shelves and about 14 days of scans. No server, no keys, no build step.
  • What it does not do yet: that count is the demo’s simulated recognition, not Gemini on a real photo. No Supabase project exists for this product, so nothing is stored, no row-level security policy runs and no edge function is called — and no accuracy figure has been measured on a real shelf.

Status

What is measured, and what is only wired

  • No accuracy figure has been measured. The harness exits 2 (NOT MEASURED) when there are no labelled photos, rather than reporting a pass. The 29-check run against a synthetic image proves the pipeline produces a number and fails a wrong label; it is a wiring proof and never accuracy evidence.
  • The ≥98% figure is a target, not a result. It is the bar a pilot agrees to hit against a customer’s own verified manual recount, and it has not been measured on a shelf.
  • The ROI numbers are the product’s own model — arithmetic on assumed hours and a wage rate, not a measured saving from any warehouse.
  • The blocking item is human. It needs real labelled shelf photos, on the order of 15–25 from one aisle including awkward and occluded ones, plus one warehouse contact who will give written permission and do the manual recount.

Under the hood

What it is built on

Stack
A React Native / Expo app and a browser dashboard over Supabase — Postgres with row-level security and six Deno edge functions. Gemini multimodal does the recognition; the drone client and QuickBooks exporter are separate integrations.
Runs where
The app starts through Expo. The dashboard above is a static page with no build step. The backend is written against a Supabase project: process-image, detect-damage, generate-audit, drone-ingest, get-image-url and submit-feedback exist as source, and none has been deployed.
Data
Warehouses, shelves, shelf_scans, discrepancies, damage_reports and utilization_history, all scoped by row-level security through an is_warehouse_member helper, with images in a private shelf-scans bucket read back through one-hour signed URLs.
Blocked on
Labelled shelf photos and a warehouse willing to be recounted. No amount of code closes that.

Ask about a warehouse

Whether it fits a warehouse you run, or what a two-week one-aisle pilot would need from you — ask before you commit anything. There is no hosted instance yet.

Open the feedback form