Building in public
Writing
Engineering deep dives on the products I'm building, and breakdowns of AI papers I've implemented — what the code actually does, not the pitch version.
Rethinking agentic RAG: what if retrieval didn't start with embeddings?
A paper breakdown and a working FastAPI build — letting the LLM write boolean expressions against an inverted index instead of reaching for a vector store first.
Verum: an outreach agent that's only allowed to send what it can prove
Eleven gates, a Sent-folder-as-truth state design, and why a guessed email address is never worth the reply rate.
ApplyWise: the AI never gets to say something I didn't tell it
Local-first storage, a schema-enforced needs_user_input, and why it never clicks Submit.
DataSync: I built the wrong product first, on purpose
An alumni portal, one timestamp column, and the dry-run flag that saved me from mailing three hundred people by accident.