- AI coding
- Codex · Claude Code · AI-assisted reviews · Development workflows
AI coding tools speed up implementation, refactoring and test writing. I treat their output like a colleague's pull request: it is read, run and tested before it is merged, and I stay responsible for every change. - AI applications
- LLM APIs · Chatbot development · AI copilots
Applications that call large language models through an API, wrapped in the product's own logic, permissions and logging, such as the DEVV copilot and the assistant on this site. - Retrieval
- RAG · Embeddings · BM25 · Vector search
Grounding answers in real content. This site's assistant uses a BM25 ranking engine over the site's own pages, and the toolkit also covers embeddings and vector search for larger knowledge bases. - AI reliability
- Retry handling · Rate limits · Fallbacks · Evaluation
Model calls fail, slow down and get rate-limited. Reliable AI features retry with limits, degrade to a working fallback and are checked against a fixed set of test questions. - AI security
- Role-aware retrieval · Authorization · Data isolation
The model is never the access-control layer. Who may see what is decided by the application before any content reaches the model, so one user's data cannot appear in another user's answer. - AI infrastructure
- Provider integration · Cost control · Logging · Observability
Integrating model providers behind a single layer, limiting usage and spend, and keeping logs that make a wrong or slow answer traceable.