Coding Rounds

Implement-from-scratch exercises that show up in AI engineering coding interviews.

The canonical set

1. Scaled dot-product attention in NumPy/PyTorch (then multi-head).
2. A minimal agent loop — model call, tool dispatch, observation, stop condition (~100 lines).
3. A RAG pipeline — chunk, embed, retrieve top-k with cosine similarity, stuff the prompt.
4. A simple eval harness — run N cases, LLM-as-judge scoring, report pass rate.
5. Streaming token handling — parse partial tool-call JSON from a stream.
6. A rate-limited async client with retries and exponential backoff.

Ground rules interviewers watch for

• Can you write the API-calling code without a framework (no LangChain crutch)?
• Do you handle malformed model output (bad JSON, refusals) gracefully?
• Do you know what to log for debugging an agent trace?

Practice against the exercises and the Build These First projects.

Work in progress

This page is a scaffold — worked examples and depth are being added. Want to help? Grab a good first issue or read the contribution guide.