Try the Demo · Quick Start · Report Bug
Industrial quality inspection is broken for small and mid-size manufacturers.
Traditional automated systems (Automated Optical Inspection - AOI) require:
- Labeled datasets curated by ML specialists - thousands of images per defect type
- Model retraining for every new product - weeks of engineering per SKU
- Physical camera and sensor calibration per assembly line
- Dedicated hardware costing $50,000-$200,000 per station
The result: 80% of manufacturing plants still rely on manual human inspection, where fatigue reduces accuracy by up to 34% after just 4 hours of continuous work. SMEs and workshops simply cannot afford automated alternatives.
The global AOI market is valued at $2.4 billion and growing at 15% annually, yet the majority of this market is locked behind enterprise-grade budgets and specialized ML teams.
AsBuilt Lens eliminates the entire training-and-calibration cycle. Instead of programming rules or labeling datasets, operators simply describe the inspection in plain English - the same way they would instruct a human inspector.
"Expected items: 4x resistor, 1x electrolytic capacitor, 1x IC chip, 1x LED"
The system handles the rest through an autonomous Dual-Agent Workflow running on the AMD MI300X:
- Describe → Write or select a natural-language inspection specification
- Capture → Upload an image or use a live camera with automatic stability detection
- [AGENT 1] Inspector → Analyzes the image against the spec to generate a structured PASS/FAIL verdict, per-item status, and bounding boxes.
- [AGENT 2] Quality Engineer → Autonomous Handoff: If Agent 1 detects a failure, Agent 2 is triggered automatically. It performs a secondary deep multimodal analysis to determine the root cause, severity, and recommends a detailed Corrective Action Plan.
Zero training. Zero datasets. Zero reconfiguration when switching products.
| Feature | Description |
|---|---|
| Zero-Shot Inspection | No ML training required. Write a spec in plain English, run the inspection. |
| Live Camera Mode | Webcam or IP camera (DroidCam/IP Webcam) with automatic motion-stability detection and hands-free capture. |
| Upload Mode | Upload any image (JPG, PNG, WEBP, BMP) for offline or batch inspection. |
| Auto-Discovery | Let the AI identify all visible components before you write your specification. |
| Structured JSON Results | Item-by-item status, detected counts, confidence scores, and human-readable summary. |
| Annotated Output Image | Color-coded bounding boxes and PASS/FAIL badge overlaid directly on the inspected image. |
| Downloadable Inspection Report | Auto-generated professional HTML report with full results - ready for QA records or audit trails. |
| Inspection History | Every run is logged in the sidebar with timestamp, pass rate, and specification. |
| Built-in Templates | Ready-to-use templates for PCB Assembly, Tool Kits, Electrical Panels, and Packaging. |
| Demo Mode | One-click demo with a pre-loaded PCB sample image - no hardware needed. |
| Metric | Value |
|---|---|
| TAM | Global AOI market ≈ $2.4 billion, growing 15% CAGR |
| Target Segment | SMEs, repair shops, quality labs underserved by traditional AOI |
| Pricing Model | SaaS - $0.05/inspection or monthly license ($99-$499/month) |
| AsBuilt Lens | Traditional AOI | Cloud AI Platforms | |
|---|---|---|---|
| Setup Time | Minutes | Weeks–Months | Days–Weeks |
| Training Data Required | None | 1000s of labeled images | 100s of examples |
| Hardware Cost | $0 (cloud inference) | $50K–$200K | $0 (but SaaS fees) |
| Product Flexibility | Change the text, done | Full recalibration | Retrain model |
| Data Sovereignty | Full (on-premise capable) | Full | Vendor-dependent |
| Analysis Type | Semantic reasoning (what + why) | Binary OK/NOK | Binary + localization |
Traditional systems give you a binary verdict: "Good" or "Bad." AsBuilt Lens reasons about what it sees - it can tell you what is missing, how many were expected versus detected, and provide factual observations about anomalies. This enables root-cause analysis directly from the inspection step.
AsBuilt-Lens/
├── app/
│ ├── app.py # Main Streamlit UI & application loop
│ ├── config.py # Environment variable loading & constants
│ ├── inspector.py # [AGENT 1] Visual inspection logic & VLM calls
│ ├── quality_engineer.py # [AGENT 2] Autonomous root-cause analysis logic
│ ├── camera.py # Webcam/IP camera management & stability detection
│ ├── utils.py # Image annotation, formatting utilities
│ ├── report.py # Automated inspection report generator
│ └── prompts/
│ └── inspection_prompt.txt # Structured prompt template for Qwen3-VL
├── assets/
│ ├── logo_dark.png # UI logo (dark mode optimized)
│ ├── demo_pcb.jpg # Demo image for one-click testing
│ ├── cover.png # Hackathon submission cover (16:9)
│ └── og-image.png # Social sharing preview image
├── docs/
│ └── technical_document.md # Full technical specification
├── .streamlit/
│ └── config.toml # Streamlit theme (Phoenix dark mode)
├── .env.example # Environment variable template
├── requirements.txt
└── README.md
AsBuilt Lens runs inference on AMD Instinct MI300X GPUs via the AMD Developer Cloud. The MI300X provides 192 GB of HBM3 memory, enabling the full Qwen3-VL-32B model to run without quantization or sharding compromises.
# Deploy Qwen3-VL on AMD MI300X via vLLM (ROCm)
python -m vllm.entrypoints.openai.api_server \
--model Qwen/Qwen3-VL-32B-Instruct \
--port 8000 \
--trust-remote-code \
--max-model-len 32768Key Configuration Details:
| Parameter | Value | Rationale |
|---|---|---|
--max-model-len |
32768 | Sufficient headroom for base64 image tokens + structured JSON output |
--temperature |
0.1 | Near-deterministic output for consistent inspections |
--max-tokens |
2048 | Enough for detailed multi-item inspection results with bounding boxes |
| Image preprocessing | 720p max, JPEG 85% quality | Optimal balance between visual detail and token budget |
- Engine: vLLM on ROCm — native AMD GPU acceleration without CUDA translation layers
- Memory Advantage: MI300X's 192 GB HBM3 fits the full 32B parameter model with ample KV-cache space for long multimodal contexts
- API Format: OpenAI-compatible REST endpoint — standard tooling, no proprietary SDKs
Performance Target: End-to-end inspection (image upload → structured result) in ≤ 8 seconds for standard PCB images at 720p resolution.
git clone https://github.com/JaDi03/AsBuilt-Lens.git
cd AsBuilt-Lenspip install -r requirements.txtcp .env.example .envEdit .env with your AMD Developer Cloud endpoint:
# AMD Developer Cloud
AMD_API_URL=http://your-mi300x-ip:8000/v1
AMD_API_KEY=your_api_key_if_needed
VLM_MODEL=Qwen/Qwen3-VL-32B-Instruct
# Camera (optional — only needed for Live Camera mode)
CAMERA_URL=http://192.168.1.100:8080/video
# Tuning
STABILITY_THRESHOLD=0.02
STABILITY_FRAMES=30
MAX_IMAGE_SIZE=720python -m streamlit run app/app.pyOpen your browser at http://localhost:8501.
Option A — Laptop Webcam
- Select "Laptop Webcam" in the sidebar.
- Click
▶️ Start Camera. - Hold your object steady. The stability bar fills up and auto-captures.
Option B — Phone as IP Camera
- Install IP Webcam (Android) or DroidCam.
- Connect your phone to the same Wi-Fi network as your PC.
- Set
CAMERA_URLin your.envto the stream URL shown in the app. - Select "Phone IP Camera" in the sidebar and click
▶️ Start Camera.
| Template | Use Case |
|---|---|
| PCB Assembly | Verify resistors, capacitors, IC chips, LEDs, oscillators |
| Packaging Verification | Check product, manual, warranty card, accessories |
| Tool Kit Inspection | Verify presence of specific hand tools |
| Electrical Panel | Check breakers, bus bars, labels, wiring safety |
| Custom Specification | Write your own natural-language checklist |
{
"inspection_passed": true,
"items": [
{ "id": "resistor", "expected_count": 4, "detected_count": 4, "status": "present", "confidence": 94 },
{ "id": "electrolytic_capacitor", "expected_count": 1, "detected_count": 1, "status": "present", "confidence": 91 },
{ "id": "ic_chip", "expected_count": 1, "detected_count": 1, "status": "present", "confidence": 97 },
{ "id": "led", "expected_count": 1, "detected_count": 0, "status": "missing", "confidence": 88, "note": "No LED visible in expected location." }
],
"summary": "3 of 4 items verified. LED component not detected — recheck position D1.",
"notes": ""
}| Component | Technology | Role |
|---|---|---|
| GPU Compute | AMD Instinct MI300X (192 GB HBM3) | Multimodal model inference |
| Inference Engine | vLLM on ROCm | High-performance serving with continuous batching |
| Vision Model | Qwen3-VL-32B-Instruct | Zero-shot multimodal reasoning |
| API Protocol | OpenAI-compatible REST | Standard integration, no proprietary SDKs |
| Frontend | Streamlit 1.30+ | Interactive inspection interface |
| Computer Vision | OpenCV 4.8+ | Camera capture, stability detection |
| Image Processing | Pillow 10+ | Image manipulation and annotation |
| Runtime | Python 3.10+ | Application logic |
| Resource | Link |
|---|---|
| 📹 Video Demo | https://youtu.be/-wU4ylbjeoE |
| 🤗 Hugging Face Space | Coming Soon |
| 📄 Technical Document | [Coming Soon] |
| 📂 GitHub Repository | github.com/JaDi03/AsBuilt-Lens |
| Phase | Feature | Description |
|---|---|---|
| ✅ MVP | Core Inspection Engine | Upload + camera modes, structured JSON results, annotated images |
| ✅ MVP | Inspection Reports | Auto-generated downloadable HTML reports for QA records |
| 🔜 Next | Hugging Face Space | Public demo for judges and community (upload mode) |
| 🔮 Future | Edge Inference | AMD Ryzen AI deployment for factory-floor use without cloud |
| 🔮 Future | Multi-Camera Networks | Multiple inspection stations feeding a central dashboard |
| 🔮 Future | ERP/MES Integration | Automatic pass/fail logging into manufacturing execution systems |
| 🔮 Future | Inspection Analytics | Defect trend monitoring and line performance dashboards |
- Event: AMD Developer Hackathon 2026
- Track: Track 3 — Vision & Multimodal AI
- Team: AsBuilt Lens Team
- Model: Qwen3-VL-32B-Instruct on AMD MI300X via ROCm + vLLM
- Challenge Alignment: Track 3 ✅ · Qwen Challenge ✅ · Ship It + Build in Public ✅
MIT License — see LICENSE for details.