Engineering · AI systems
AI systems
3 write-ups on ai systems.
- Keeping embeddings fresh without blocking a requestSemantic search is only as good as the freshness of its embeddings. An indexing pipeline with one authoritative write path, batch backfill, real-time incremental updates, debouncing and dead-letter handling.
- Getting a data structure out of a language modelA tool-calling ReAct agent that turns unstructured documents and third-party profile data into a validated record — and what breaks when you treat model output as trustworthy.
- Asking a model to score a match, and storing the reasonsGenerating a 0–100 relevance score for an application, with the specific evidence behind it — why the explanation matters more than the number, and where scoring quietly goes wrong.