# Builderr starter brief: restaurant kitchen CCTV monitor

## The job

Build a tool for a restaurant or cloud kitchen that can answer practical questions from long fixed-camera footage. Examples: did staff wear caps, when did a handoff happen, and what did the camera fail to show?

## What the agent receives

- Public fixed-view kitchen videos, including long clips.
- A JSON file containing questions about those videos.

## What it must return

- One answer for every question.
- Evidence: a timestamp or frame span supporting the answer.
- `not_visible` when the footage does not show enough. Guessing is worse than admitting uncertainty.
- A run log with frames inspected, model calls, runtime, and estimated cost.

## What matters

1. Answer the question accurately, including counts, order of events, and timestamps.
2. Stay under the hard cap of $0.30 estimated model/API cost per 60 minutes of video.
3. Finish within the stated runtime and frame limits.
4. Make the run repeatable with one clear command.

## A beginner-friendly way to start

1. Read the sample videos and sample questions on the [challenge page](https://builderr.ai/kitchen-video).
2. Ask an AI coding assistant to create a small `answer.py` that reads the video and question JSON and writes answers plus a log.
3. Start with one question type, such as “was a cap visible?”, before supporting every question.
4. Sample the video cheaply first, inspect only likely moments, and save the evidence timestamps.

You do not need to train a vision model. You do need to show the answer, the evidence, and the cost. Open-source/local models are recommended; cloud models are allowed if the run stays within the same cap.

## Full truth

The [challenge page](https://builderr.ai/kitchen-video) contains the public videos, sample questions, hidden-question policy, scoring weights, cost cap, and submission format. The full technical spec is also available [here](https://builderr.ai/kitchen-video-challenge-draft.md).