1 · Developer setup agent (team build)
Problem: a new developer had to clone dozens of repos and hand-configure MongoDB, Redis and other dependencies. Onboarding took a long time and broke often.
Outcome: now used by all developers on the team. It's an agent with real tools and a validation loop, not a chat wrapper.
2 · Engineering knowledge RAG (built by me)
Internal engineering docs are chunked and embedded into Pinecone, and an LLM answers from the retrieved context. The goal was fewer repeated "where is this documented?" questions and faster onboarding. Architecture details are in RAG Architecture.
3 · LLM-assisted content migration (built by me)
The key design choice: the LLM proposes and deterministic code validates. The model infers content types and fields; code enforces the target schema and emits the import format. The fuzzy part gets AI, and correctness stays in code.
What ties them together
- Treat AI features as systems: tools, validation, retrieval, failure handling.
- Put AI where the work is fuzzy (understanding a site, interpreting logs) and keep code where the work must be exact (schemas, provisioning).
- Measure adoption. The setup agent spread across the team because it worked.