In an AI engineering interview, three real projects beat ten certificates. The right projects prove you can design, build, and reason about systems — not just follow a tutorial. Here are five that stand out, and exactly what each one signals.
1. Production RAG service
Ingestion → chunking → embeddings → retrieval → reranking → grounded answers with citations, plus a small eval set. Signals: you can ground models in data and prove quality. Start with the step-by-step build.
2. Tool-using agent
An LLM that calls tools, keeps memory, and knows when to stop, with guardrails. Signals: you understand agentic patterns and safety — and when a workflow is better.
3. Chat-with-your-docs app
A full-stack app on top of your RAG service, deployed and observable. Signals: you ship end to end, not just notebooks.
4. Eval harness
A small framework that scores prompt/model changes against a test set. Signals: you make decisions with evidence — the top-1% habit most candidates lack.
5. Domain assistant
A vertical agent for a specific workflow, built on synthetic data. Signals: you can translate a real problem into a reliable AI system.
Make them legible
Each project needs a clean README, an architecture diagram, a short demo, and a "how I'd discuss this in an interview" note. That packaging is what turns a repo into an offer.
These map directly to the AI Engineer Roadmap. Pick your starting point on the learn page, then rehearse with the interview questions.