Ph.D. in Engineering Physics building AI agents that act on real software, with the guardrails in code and in the database rather than in the prompt.
Before that, four years of machine learning on noisy sensor signals (battery virtual sensing, fibre-optic instrumentation).
- db-enforced-agent-permissions: Postgres RLS as the boundary for an LLM agent, plus a LangGraph tool-calling agent and evals measured on 40 synthetic conversations: a leaked order id gave 2 breaches without guardrails and 0 after a per-conversation scope enforced in the database.
- mercadolivre-mcp: MCP server for the Mercado Livre seller API, where every write returns a preview and runs only with explicit confirmation, covered by 196 tests against a fake API.
- timelapse-studio: local tool that turns the photo sequence of a 3D print into a timelapse video, with colour correction and stabilization.
- probabilistic-load-forecast: day-ahead probabilistic load forecasting on ONS open data, validated walk-forward.
- fbg-fp-demod: peak-tracking demodulation for Fabry-Perot and FBG fibre-optic sensors, a Python port of the MATLAB pipeline, cross-checked against it on recorded battery-cell spectra.
- rag-chatbot: FastAPI retrieval-augmented generation API over a Markdown knowledge base, with conversations stored in PostgreSQL.