AI-powered equipment inspection platform. Stream video from any camera to cloud AI, get real-time visual analysis on a live dashboard with full inspection reports and fleet memory.

The SDK connects a camera (webcam, drone, Raspberry Pi, RTSP) to a GPU backend running Qwen2-VL on Modal. An operator watches the live feed on a web dashboard with AI analysis, zone tracking, voice Q&A, and a generated PDF report.

Camera (SDK)  --WebSocket-->  Modal Backend (Qwen2-VL)  --WebSocket-->  Dashboard

1. Sign in at the CatWatch dashboard with your email (magic link)
2. Go to API Keys > Create Key and copy the key (looks like cw_live_xxxxxxxxxxxx)
3. Install the SDK:

pip install catwatch

4. Run:

from catwatch import CatWatch

cw = CatWatch(api_key="cw_live_YOUR_KEY")
cw.connect(source=0, mode="797")
print(f"Dashboard: {cw.dashboard_url}")
cw.run()

Open the printed dashboard URL in a browser to see the live inspection.

"general" — general safety and scene monitoring. The VLM flags hazards, unsafe conditions, and person-related risks.

"797" — CAT 797F defect inspection. The VLM inspects for mechanical defects, wear, leaks, and structural damage specific to the Caterpillar 797F mining truck. It auto-identifies equipment type and builds a dynamic zone checklist. Reports "No 797F visible" only when the truck is genuinely absent from frame.

Only one inspection runs at a time per user. Starting a new session automatically closes any previous one.

source can be:

• 0 — USB webcam or built-in camera
• "/path/to/video.mp4" — pre-recorded video file
• "rtsp://192.168.1.10:8554/cam" — RTSP / IP camera stream
• A Picamera2 object — Raspberry Pi CSI camera

Pass identity so the AI remembers past inspections of the same unit or site:

# 797 mode — CAT 797F defect inspection
cw.connect(source=0, mode="797", unit_serial="797F-001", model="CAT 797F", fleet_tag="fleet_alpha")

# General mode — site name
cw.connect(source=0, mode="general", location="warehouse-7B")

If you don't pass unit_serial or location, the dashboard falls back to browser geolocation so inspections at the same physical location share memory automatically.

On later inspections of the same unit or site, the dashboard shows prior findings, and the VLM analysis is primed with historical context. Fleet-wide patterns are surfaced across units with the same fleet_tag.

Register before calling cw.run():

@cw.on_analysis
def handle_analysis(msg):
    data = msg["data"]
    print(f"[{data['severity']}] {data['description']}")

@cw.on_finding
def handle_finding(msg):
    f = msg["data"]
    print(f"[{f['rating']}] {f['zone']} — {f['description']}")

@cw.on_session_ended
def handle_ended(msg):
    print("Session ended by operator")

cw.run()

Available callbacks: on_analysis (VLM, every ~3s), on_finding (persisted finding), on_voice_answer (voice Q&A response), on_report (generated report), on_viewer_joined, on_mode_change, on_session_ended.

# Ask about the current frame
cw.ask("Is there hydraulic fluid leaking?")

# Generate an inspection report
cw.generate_report(model="CAT 797F", serial="797F-001")

# Switch mode mid-session
cw.set_mode("general")

# Send raw payload
cw.send({"type": "voice_question", "text": "What's the condition of the tracks?"})

Once the SDK is streaming, the dashboard shows:

• Live video feed with real-time VLM analysis overlaid
• VLM analysis panel — severity, description, confidence, with severity trend tracking across repeated analyses of the same zone
• AI Insights — cross-signal correlations and pattern detection
• Zone checklist — in 797 mode, the VLM identifies equipment and builds a dynamic checklist. In general mode, zones appear as they're discovered
• Voice Q&A — mic button on the video feed to ask questions about what the camera sees
• Unit/Site Memory — history from prior inspections of the same unit or location
• End Inspection — stops the session and shuts down the SDK connection

After an inspection, view the full report (PDF download), all findings (filterable by severity), 3D inspection view, fleet-wide similar findings, and an AI-generated insurance claim document.

sudo apt update && sudo apt install -y python3-pip python3-venv libcap-dev
python3 -m venv ~/catwatch-env && source ~/catwatch-env/bin/activate
pip install catwatch[picamera]

from picamera2 import Picamera2
from catwatch import CatWatch

cam = Picamera2()
cam.configure(cam.create_video_configuration(main={"size": (1280, 720)}))
cam.start()

cw = CatWatch(api_key="cw_live_YOUR_KEY")
cw.connect(source=cam, mode="797", unit_serial="DRONE-PI-001")
cw.run()

To run on boot, create a systemd service at /etc/systemd/system/catwatch.service:

[Unit]
Description=CatWatch Inspection
After=network-online.target
Wants=network-online.target

[Service]
User=pi
WorkingDirectory=/home/pi
ExecStart=/home/pi/catwatch-env/bin/python /home/pi/inspect.py
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

sudo systemctl enable catwatch && sudo systemctl start catwatch

• SDK: Python — OpenCV, websockets, numpy
• Backend: Modal (A100 for Qwen2-VL + Whisper), FastAPI orchestrator
• Frontend: Next.js 15 on Vercel, PostgreSQL via Prisma, Magic.link auth
• Memory: Supermemory for cross-inspection knowledge
• Insurance Claims: OpenAI GPT-4o-mini for AI-generated claim documents

See DEV.md for the developer guide (deploying, environment setup, architecture details).

• ModuleNotFoundError: catwatch — run pip install catwatch
• ValueError: mode must be one of general, 797 — update your SDK to v0.3+ and use mode="797" (not mode="cat")
• "Invalid API key" — create a key in Dashboard > API Keys
• No VLM analysis — the dashboard must be open (GPU only runs when a viewer is connected). VLM fires on a cadence; give it ~5 seconds after the dashboard loads
• Dashboard video blank — check that NEXT_PUBLIC_BACKEND_WS matches your Modal deployment URL
• Supermemory empty on second run — make sure SUPERMEMORY_API_KEY is set in your web .env.local
• Insurance claim button missing — only shown for completed (non-active) sessions with findings. Requires OPENAI_API_KEY in .env.local