Start a project
AI intelligence system

StockReconcile

From box counting to continuous reconciliation.

A vision system that watches warehouse activity and continuously compares what is physically on the shelf against what the WMS believes, so discrepancies surface within hours instead of at the next stock count.

Warehousing Retail Supply Chain Computer Vision WMS integration
Sector
Warehousing, Retail & Supply Chain
Year
2025
Duration
7 months
Team
6 people
Platforms
Edge · Web console
Status
In production
StockReconcile — visual identity for the case study
Hours Discrepancy detection Previously a quarterly cycle
Clip-backed Every discrepancy With the transaction it should match
Continuous Reconciliation cadence Against the WMS stream
No shutdown Operational impact The floor keeps running

01 — The problem

Physical inventory counts happened quarterly, shut down half the floor for two days, and produced a number that was already stale when it was published. Between counts, the WMS was assumed correct because nothing contradicted it.

The discrepancies were not random. Misplacement, mis-picks and unrecorded movements accumulated in predictable places, but by the time a count found them the trail was three months cold and no cause could be established.

02 — What we did

We stopped treating counting as an event. Cameras already covering the aisles observe pallet and case movement continuously, and every observation is reconciled against the WMS transaction stream as it happens.

The output is deliberately narrow: a discrepancy queue. Rather than claiming a perfect real-time count, the system reports specific locations where physical observation and system state disagree, with the clip and the transaction that should have matched.

A quarterly count tells you that you were wrong. It never tells you when you became wrong.
Interface

What it looks like in use

Two of the screens that carry the most weight in daily use, rebuilt here from the production design system.

stockreconcile.aurezalabs.com/queue
Discrepancies · 24h9 OPEN
Bin D-18 · movement, no transactionCLIP 04:12
Bin A-03 · transaction, no movementCLIP 11:38
Bin F-22 · occluded viewLOW CONF
Directed counts issued4
Hours, not quarters

A narrow, honest outputRather than claiming a perfect live count, the system reports specific locations where physical observation and system state disagree.

stockreconcile.aurezalabs.com/accuracy
Inventory accuracyCONTINUOUS
0SHUTDOWN DAYS
4hMEDIAN DETECTION
Weekly accuracy

Accuracy as a live measureInventory accuracy is tracked continuously instead of appearing as a quarterly snapshot that is stale on publication.

Capabilities

What the system does

Grouped by the job each set of capabilities exists to do, rather than by which team built it.

Observation

01 · Watching the aisles

Pallet and SKU recognition

Case and pallet identification from label, shape and placement context across aisle cameras.

Multi-camera tracking

Movement followed across camera boundaries so a pallet leaving one view is the same pallet entering the next.

Location inference

Physical position resolved to the bin and level of the racking model rather than to a camera frame.

Occlusion handling

Confidence lowered rather than fabricated when a view is blocked, with the gap reported.

Reconciliation

02 · Physical against digital

WMS integration

Live transaction stream consumed and matched against observed movement within a tolerance window.

Physical-versus-digital comparison

Continuous comparison producing a discrepancy queue rather than a claimed absolute count.

Misplacement detection

Stock present in a location the system does not expect, flagged with both records.

Unauthorised movement alerts

Movement with no corresponding transaction raised while the clip is still available.

Operations

03 · Acting on it

Discrepancy triage

A worklist ordered by value and age, each item opening onto its clip and expected transaction.

Targeted counts

Cycle counts directed at locations the system flags instead of scheduled blindly by zone.

Root-cause patterns

Discrepancy clustering by shift, zone and equipment to separate process problems from one-offs.

Accuracy trending

Inventory accuracy tracked continuously as a live measure rather than a quarterly snapshot.

Architecture

How it is put together

Layer by layer, with the reason each one exists — because the reason is usually the interesting part.

01Edge vision

Aisle cameras with on-site inference producing movement events; footage is retained briefly and only around flagged events.

  • Edge GPU
  • Detection
  • Short retention
02Spatial model

Racking model mapping camera coordinates to bin and level so observations land in warehouse terms.

  • Calibration
  • Racking model
03Event matching

Temporal matching of observed movements against WMS transactions with configurable tolerance and confidence.

  • Stream matching
  • Tolerance windows
04Discrepancy service

Queue of unmatched observations and unobserved transactions with evidence attached to each.

  • PostgreSQL
  • Evidence linking
05Console

Triage workspace with clip playback, expected transaction and directed count generation.

  • React
  • Video playback
  • WMS API

Technology

Edge

  • Edge GPU
  • TensorRT
  • GStreamer

AI & Vision

  • PyTorch
  • Multi-object tracking
  • Label recognition

Backend

  • Python
  • FastAPI
  • Kafka
  • PostgreSQL

Integration

  • WMS API
  • ERP sync
  • Webhooks
Outcome

What changed

Measured against how the operation ran before, not against a benchmark chosen after the fact.

  • Discrepancies surface in hours. The trail is still warm, so causes can actually be established.

  • Counts became targeted. Cycle counting goes where the system flags rather than sweeping zones on a schedule.

  • The floor stopped shutting down. Continuous reconciliation replaced the two-day quarterly count.

  • The system states its uncertainty. Occluded views lower confidence and report a gap rather than inventing a number.

How it ran

  1. Months 1-2

    Coverage and calibration

    Assessed existing camera coverage and built the racking model that observations map into.

  2. Months 3-4

    Recognition and tracking

    Pallet and case recognition with cross-camera tracking through occlusion.

  3. Month 5

    WMS reconciliation

    Transaction stream integration and the tolerance-window matching engine.

  4. Months 6-7

    Triage console

    Discrepancy queue, clip playback, directed counts and accuracy trending.

Related work

From the same practice

All fifteen projects

Manufacturing & Production Quality

LineTrace

From defect detection to defect causality.

A multimodal system that correlates line video, machine telemetry, batch records and operator actions to explain why a defect happened, not merely that it did..

2025 · 10 months
Telecommunications & Network Operations

NetCause

Not alarm detection. Failure hypothesis reduction.

A correlation engine that turns an alarm storm into a short ranked list of probable root causes by reasoning over network topology and fault propagation rather than alarm counts..

2025 · 8 months
Oil & Gas · Asset Integrity

PipeIntegrity

From a single inspection to longitudinal asset integrity.

A vision system that detects and localises pipeline deterioration, then tracks each defect across inspection cycles so maintenance is prioritised by progression rather than by appearance..

2024 · 9 months
Something similar?

Running into the same problem StockReconcile solved?

Tell us what you run. We will reply within two business days with what we would build for your situation — and, just as usefully, what we would leave out.

Start a conversation More work