About
I build AI systems where the model is one component among checks. That means validating inputs before a model acts, measuring behavior against recorded cases, and putting guards around state changes.
The four case studies here are prototypes. Each one states what was measured, what was not, and what part of the work is mine. Where an assessment brief, a bulk commit or an AI-assisted workflow is part of the history, the page says so.
Current focus: LLM pipelines and retrieval with evaluation, structured extraction from noisy inputs, and the backends around them.
Technology index
Technologies as used in each project. This is not a rating of skill; depth is shown by the case studies.
- DuckDB
- Retrieval over research papers with verify and repair
- Express
- Ride-pooling lifecycle with guarded transitions
- FastAPI
- Retrieval over research papers with verify and repair, Speech and report extraction that does not invent values
- faster-whisper
- Speech and report extraction that does not invent values
- LLM provider APIs
- Two-stage email drafting pipeline
- Next.js
- Ride-pooling lifecycle with guarded transitions
- PostgreSQL
- Ride-pooling lifecycle with guarded transitions
- Pydantic
- Two-stage email drafting pipeline
- Python
- Two-stage email drafting pipeline, Retrieval over research papers with verify and repair, Speech and report extraction that does not invent values
- sentence-transformers
- Retrieval over research papers with verify and repair
- Streamlit
- Two-stage email drafting pipeline
- Tesseract
- Speech and report extraction that does not invent values
- TypeScript
- Ride-pooling lifecycle with guarded transitions