Work
Visual reconciliation for returned kits
A working prototype: a vision model reconciles a returned kit against its manifest item by item, with a bounding box, a confidence score and a condition class for each item, plus a replenishment recommendation.
- What it proves
- Computer vision for counting and reconciliation, where the audit trail matters as much as the answer.
- Method stage
- Built from
- Vision-model structured extraction
- Bounding-box item detection
- Quantity counting
- Per-item confidence scoring
- Condition classification
- Auto-generated replenishment recommendation
The problem
Checking a returned kit against its manifest is repetitive work, and a wrong count has consequences. A missing or contaminated item in a kit that goes back into service is a safety gap. Most automated counts give a total and no way to check how it was reached.
What we built
The system takes a photograph of a returned kit and returns structured output. For each item it detects, it gives:
- a bounding box showing where in the image the item was found
- a count where several are expected
- a condition class: usable, damaged or contaminated
- a confidence score for each of those judgements
Where the manifest and the detected contents disagree, or confidence on an item is low, the system writes a replenishment recommendation.
The system never collapses the result into a single pass or fail, because a verdict of “kit incomplete” cannot be audited. A reviewer can check each item: its name, its location in the image, its class and the confidence.
What it proves for a client
Per-item boxes and confidence scores let a reviewer confirm, override or reject each judgement, so an audit can stand behind the count.
Method stages
It sits at Stage 02 · Prove the value, where the evaluation is written before the prototype and measured against a manual baseline. See the KARPINITY Method.