AI Engineer building reliable Python systems for Applied AI.

A macOS assistant that combines selected screen context, microphone input, meeting audio, local OCR, relevance-ranked retrieval, transcription and multiple AI providers.

My work

Built focused context selection, provider-aware routing, consent controls, secure Electron boundaries, automated tests and release checks.

Implemented controls

User-selected inputs, consent-led capture, sandboxed execution, context isolation and provider-aware routing.

A local-first review-intelligence system for adaptable CSV datasets, semantic retrieval and grounded answers with inspectable citations. Citation validation and a single bounded repair pass keep responses grounded and evidence-linked.

My work

Built adaptable field mapping, stable record identity, recoverable vector indexes, citation validation, bounded repair, privacy-safe diagnostics and an installable CLI.

Evaluation result

Clean exact-commit 30-case benchmark: Semantic Recall@5 0.913 versus 0.770 for BM25; answer success and citation validity 0.880.

A leakage-aware football forecasting pipeline with chronological train, calibration and frozen test windows, using calibrated XGBoost and Poisson probabilities.

My work

Built the data pipeline, temporal features, model evaluation, FastAPI service and Streamlit interface, with dedicated workflows for live, retrospective and synthetic datasets.

Evaluation result

Evaluated across 1,140 rolling-origin test matches with 53.77% accuracy, using a 56.70% bookmaker benchmark for comparison.