Restaurant ops agent

Live: https://restaurant-ops-agent.vercel.app

An AI agent that answers a restaurant owner's questions about their own sales and their own written policies, and proposes actions a person approves before anything happens. The model writes SQL, searches policy and drafts actions. Code checks every query as a parse tree, limits it to one restaurant, runs it read-only, and records every step with its cost. Nothing changes without a person.

Two fictional restaurants, generated orders, no real sales data. Built as an applied AI case study. The site walks through it:

• Scenarios. Five questions, simplest first, each with what should happen written down before it was asked, the captured answer, and the full trace. Each can be rerun live.
• Evaluation. 52 questions with a known right answer, graded by code, every failure shown in full, and the same cases on other models, efforts and retrieval thresholds.
• Architecture. The request path as a diagram, one section per step with a link to the file, the decisions, and what it does not do.

Layout

api/     FastAPI. The agent loop, the SQL guard, policy retrieval, proposal
         checks, the evaluation harness and its cases and runs.
web/     Next.js. The site. Deploys to Vercel.
data/    Schema, restaurant profiles, the dataset generator, the policy
         documents.
infra/   Postgres init for docker compose.

Run it

Docker runs the database and the API. The site runs with pnpm, as it does on Vercel.

cp .env.example .env                         # add ANTHROPIC_API_KEY and OPENAI_API_KEY
docker compose up -d --build --wait          # Postgres (loaded from data/out) and the API on :8000
docker compose run --rm tools sh -c 'python -m app.policies.index --dsn "$DATABASE_URL" --dsn "$EVAL_DATABASE_URL"'
cd web && pnpm install && pnpm dev           # the site on :3000

Everything else runs in the same image against this checkout:

docker compose run --rm tools python -m pytest                       # the API tests, database included
docker compose run --rm tools python -m app.evals.run                # the evaluation set, against the frozen database
docker compose run --rm tools python ../data/generate.py --out ../data/out --profiles ../data/profiles --months 24 --seed 42

Model: Claude Sonnet 5.5. Embeddings: OpenAI text-embedding-3-small. Database: Postgres with pgvector.

Licence

MIT.