Projects With Production-Shaped Problems

Problem

Job-search activity is fragmented across inboxes, notes, spreadsheets, and manual reminders.

What I Built

Built a Gmail-backed recruiting inbox, model trace surface, review gates, career memory, application tracking, and deployment-ready product flow.

Impact

Turns ad hoc recruiting email into auditable application state, reminders, and follow-up execution without letting model output mutate state unchecked.

Problem

A public hackathon artifact needs to prove the product loop without exposing private Gmail data, secrets, or the full hosted codebase.

What I Built

Packaged a credential-free judge demo with sanitized fixtures, optional Gemma model path, deterministic fallback, model traces, and 15-case pipeline evaluation.

Impact

Makes the recruiting-agent system inspectable: evidence becomes extracted workflow state only when confidence and review gates allow it.

Problem

Chinese Chess needs correct legal move generation, compact board encoding, and measurable training loops before model strength matters.

What I Built

Built board/FEN logic, legal policy masking, residual CNN policy-value heads, PUCT-style MCTS, self-play, supervised training, and smoke evaluation presets.

Impact

Shows deep technical work beyond app UI: rules correctness, model optimization, reproducible configs, and measurable agent-style search behavior.

Problem

Small service teams need dependable scheduling without expensive, overbuilt booking software.

What I Built

Building a clean booking product surface with mobile-friendly appointment flows, follow-up automation, shared availability, and multi-language support.

Impact

Currently an unfinished product that shows the direction: faster setup, simpler booking operations, and clearer scheduling visibility for lean teams.