I build things, learn difficult stuff, and explore better ways to {think}.

What was hard

Faculty inboxes need a way to separate serious outreach from spam without silently losing a candidate who deserves review.

What I built

A Gatekeeper routes uncertain cases to a reviewable queue, then five agents debate each candidate over a shared transcript. The final verdict includes receipts from the professor's publications, and no email sends without human approval.

What I learned

An agentic workflow needs an audit trail. A verdict is more useful when a professor can replay the reasoning and decide what happens next.

What was hard

Graduate-school guidance has to keep a student's context intact while making program recommendations they can inspect and act on.

What I built

A session-aware multi-agent system with a PersistentChatAgent wrapper, human-in-the-loop checkpoints, Elastic MCP program search, and a FastAPI and Next.js application stack.

What I learned

Durable sessions and explicit checkpoints make an agent easier to trust when the decision has real consequences for the person using it.

What was hard

A real-time tutor is only useful when its voice and vision session survives the messy parts of a live connection.

What I built

A Gemini Live API tutor on FastAPI and Cloud Run, with WebSocket streaming, structured error envelopes, and a dismissible recovery banner.

What I learned

Realtime reliability is state management: reconnects, audio ownership, and session hydration are product behavior, not edge cases.

What was hard

Academic answers need evidence, not plausible language. The hard part was making retrieval quality visible and measurable.

What I built

Fine-tuned BAAI/bge-large embeddings over the ACL Anthology and used vector retrieval to ground scholarly answers.

What I learned

Retrieval needs its own evaluation. Measuring the retrieval layer made the improvement against baseline LLM answers concrete.