The platform

Built for the factory floor,
not the data science lab

Five pages. One workflow. Every annotation tracked, every decision auditable, every QA reviewer supported by AI — not replaced by it.

Five QA pages

Everything your team needs,
nothing they don't

📊

Dashboard

Live metrics: total annotations, auto-approved count, pending queue size, auto-approval rate, and defect class distribution chart. Know your line status at a glance.

🔍

Annotate image

Upload any image. Get confidence score, routing decision, and full probability breakdown across all 6 defect classes in under 10 seconds.

📋

Review queue

Human reviewer interface for medium-confidence items. Approve or reject with label correction. Every decision logged to the audit trail.

🤖

AI assistant

Claude-powered chat built directly into the platform. Ask about defect types, QA decisions, manufacturing processes — in plain language.

📜

Audit trail

Full timestamped log of every AI annotation and human review action. Complete traceability for warranty disputes, recalls, and compliance.

⚙️

Configurable thresholds

Auto-approve and auto-reject thresholds are environment variables — no code changes needed to tune the routing engine per deployment.

Technology stack

Production-grade from day one

AI MODEL
SegFormer (nvidia/mit-b0)
Fine-tuned on NEU Steel Surface Defect dataset · 100% Gate 1 accuracy
BACKEND API
FastAPI (Python)
6 REST endpoints · SQLite audit trail · confidence routing engine
FRONTEND
Streamlit → Next.js (in progress)
5-page QA reviewer UI · Claude AI assistant tab
CONTAINERIZATION
Docker
Single-container deployment · platform-independent
CLOUD
Azure + HuggingFace Spaces
$1K Azure credits active · live public demo URL
AI ASSISTANT
Anthropic Claude API
Embedded QA chat assistant · factory-context prompting

See it running on real defect data

The live demo runs on HuggingFace — no install, no login. Upload a steel surface image and see the routing decision in real time.