Watch: Context for Reliable Agents
Live demo · Drone roof inspection

Inspect drone footage on Chalk GPUs

This demo shows how to run Qwen3-VL-8B on a Chalk GPU to identify visible roof damage and draft an inspection summary. Choose a flight to try it, then explore the code to see how it works.

See the DemoTry Chalk

Check each roof finding against the source frame

A vision model summarizes drone footage. Use the timestamps to inspect its evidence.

Try this experiment · expected results

  1. 1Analyze a sample flight

    Choose an included flight and select Inspect roof.

    Expected: Findings arrive frame by frame before the overall summary.

  2. 2Verify a finding

    Select a finding to jump to its timestamp and compare the description with the visible roof.

    Expected: The footage lets you judge whether the finding is supported; the summary is decision support.

qwen3-vl-8bOpen in new tab

Inspect the sample footage, then check each finding against its video frame. Chalk serves the vision model; the output is evidence to inspect, not a damage guarantee.

Sample footage: U.S. Army Corps of Engineers drone video from Operation Blue Roof after Hurricanes Ian and Ida, in the public domain via DVIDS. The summary is decision support, not a coverage determination.

Built on Chalk

Chalk serves the vision model on an NVIDIA L40S and runs the inspection app on CPUs. The app samples keyframes, requests findings, and assembles a summary linked to the footage.

  • Serve an open vision model

    Deploy Qwen3-VL-8B with the stock vLLM container. Chalk provisions the GPU and checks that the model is ready before routing requests to it.

  • Authenticated model calls

    The model endpoint requires authentication. The inspection app uses its Chalk workload identity to call it from a separate CPU service.

  • Stream findings as they arrive

    The inspection app streams each completed frame’s findings through its Chalk endpoint. The browser can show those results while the remaining frames and summary are still being processed.

  • Structured findings for your workflow

    The app supplies JSON schemas to vLLM for damage types, severities, and bounding boxes. It links the model’s findings to evidence frames for an adjuster to review.

Deploy your own

Run it on your own

Clone the example, log in to Chalk, and run the deploy script. It starts vLLM with Qwen3-VL-8B on an NVIDIA L40S, deploys the CPU inspection app, and prints the demo URL.

You will need uv, the Chalk CLI, and a Chalk environment with L40S capacity. Adapt the app’s damage categories and summary schema for your inspection workflow, then connect its streaming API to your intake process.

chalk-ai/roof-inspection
git clone https://github.com/chalk-ai/roof-inspection
cd roof-inspection

# Or export CHALK_CLIENT_ID and CHALK_CLIENT_SECRET.
chalk login

# Deploys the vision-language model on a GPU and the inspection app, prints your demo's URL.
./deploy.py

Bring vision models to roof inspections

Turn inspection footage into findings an adjuster can review. Talk to an engineer about serving vision models alongside your claims workflow on Chalk.

Talk to an engineerExplore model serving