amtrak-explorer

Browse Amtrak stations and routes, and ask plain-English questions about schedules ("first train from New York to Boston next Thursday?", "Friday out, Sunday back from Philadelphia"). Built on real Amtrak GTFS data. A learning project: SQL, analytics instrumentation and AI-product evals.

Live: https://amtrak-explorer.vercel.app (behind Vercel SSO, so you need to be signed in)

Stack

Layer	What
Database	Neon (serverless Postgres)
Frontend	Plain HTML/JS/CSS, no framework, no build step
Backend	Vercel serverless functions (api/); all database access goes through them
AI	Anthropic API, one pinned model, three fixed-SQL tools (the model never writes SQL)
Hosting	Vercel, auto-deploys on push to main

Run it locally

npm install
npm test        # unit tests, free, no network

The api/ functions need real credentials, so test them on the deployed site or with vercel dev. npx serve only serves the static page.

Run the evals

The eval harness runs 40 hand-written questions through the real ask loop and appends a row to eval/RESULTS.md. It makes real model calls (a few tens of cents), so run it on demand, never in CI.

vercel env pull .env.local    # DATABASE_URL + ANTHROPIC_API_KEY (git-ignored)
npm run eval                  # add -- --no-record to skip writing a row

• If keys are missing, try npx vercel env pull .env.local --environment=production.
• A Sensitive variable pulls as an empty value or placeholder: paste your own key into .env.local.

Where to read more

• CLAUDE.md: how the project works, the architecture rules, the AI layer, data gotchas, roadmap and open items. Start here.
• docs/PRD.md: what this phase set out to build and why.
• eval/RESULTS.md: the eval run log.