Projects you can run, test and push

The final project. One AI feature that actually works, wrapped in the things that make it a product: API-key auth, input validation, per-customer daily limits, token and cost tracking, a clean browser UI, tests and a deploy file. It runs in demo mode with no key, so you can build and test the whole flow before paying for a model.

A small but genuine React + Vite app. The data model lives in one pure module with seven tests, so you can change the UI without breaking the logic. It saves tasks in localStorage, works on phones, and builds to static files you can put on any host.

npm test → 7 tests passing

A dependency-free retrieval-augmented generation pipeline you can read in one sitting. It chunks documents, builds deterministic embeddings, ranks passages with a hybrid of exact terms and cosine similarity, and answers with citations or an honest "I don't know".

pytest -q → 5 passed

A minimal agent loop you can actually reason about. The model is injected as a plain decide() function, so the control flow is deterministic and testable without an API key. Every action and observation is recorded in a trace, and a step budget stops runaway loops.

pytest -q → 3 passed

A tiny evaluation harness that treats prompt changes like code changes. Load cases from JSONL, run your model, check required and forbidden text, print failures and exit non-zero in CI. The demo model scores 4/4; break it and watch the score fall.

pytest -q → 5 passed