All projects · AI Agents & RAG
StockVision: AI Visual Inventory Counter
An enterprise-grade visual inventory intelligence platform that localizes, counts, and tracks stock items across warehouse shelves and retail…
An enterprise-grade visual inventory intelligence platform that localizes, counts, and tracks stock items across warehouse shelves and retail displays. Features structured object detection with exact bounding box coordinates (x, y, w, h), confidence scores, instance aggregation, low-confidence review alerts, real-time par-level stock reconciliation against expected inventory, temporal difference comparisons across historical scans, structured AI difference narrative synthesis, multi-format export (CSV, JSON, PDF), and a WebRTC live camera scanner with multi-frame IoU and centroid spatial-temporal tracking to prevent duplicate counting during camera pans (The Clue Mandate).
- Structured Localization (The Clue Mandate): Replaces vague textual image descriptions with strict structured telemetry: exact bounding box coordinates (x, y, w, h), item classification, confidence scores, and instance counts
- Par-Level Expected Stock Reconciliation: Compares detected quantities against expected ERP stock levels, automatically flagging missing (-N), excess (+N), and unexpected items with color-coded status badges
- Interactive Human Review & Correction Lifecycle: Enables operators to correct misidentified objects, delete false detections, and manually draw bounding boxes over missed items on an interactive canvas
- Low-Confidence Verification Alerts: Automatically highlights low-confidence detections (<0.90) with dashed warning boundaries and prompts human review before final audit lock
- Temporal Scan Comparison & Ghost Diffing: Compares scans of the same location across different timestamps (e.g. 10/03 vs 12/03), calculating appeared, disappeared, and quantity-adjusted items with AI narrative difference synthesis
- Bonus Live Camera Tracking & Anti-Duplicate Protection: Features live WebRTC camera mode with IoU and centroid distance tracking across consecutive video frames to prevent recounting the same physical object during camera sweeps
- Comprehensive Export Suite: Instant RFC-4180 CSV export with coordinates, structured JSON schema with audit trail, and executive printable PDF report
Built with: HTML5 / Canvas, SVG Overlays, Computer Vision (Bounding Boxes), Spatial IoU Tracking, Centroid Distance Matching, WebRTC Video Stream, Vanilla JavaScript, CSS3 Glassmorphism, Python 3.12 Engine, CSV / JSON / PDF Export