Work

Self-hosted Jev decision readout.

Noul, Score and Choice readouts from one forward pass, served with vLLM and PagedAttention on Kubernetes.

longmem: memory for AI assistants.

Open-source cross-project memory so important decisions aren't lost. Hybrid semantic + keyword search, MCP server, published on PyPI.

Observability platform for ML workloads.

Replaced Grafana–Prometheus with OpenTelemetry, ClickHouse and SigNoz, with alerting to Teams and Slack. Incident response 30% faster.

Modernized a production ML platform.

End-to-end migration: Airflow → Prefect, plus a Rust-based Python toolchain (Poetry → uv, pandas → Polars, Loguru → Logly). Zero downtime.

Unified API layer.

One backend for MDPI's internal and external services (FastAPI, PostgreSQL), with versioned Parquet backups on S3.

From research to hospitals.

A co-registration algorithm 40× faster on CPU, then 4 research prototypes in production at hospitals across the Boston area.

CarGo Relay: parcels on everyday routes.

C2C parcel delivery matched to the routes people already travel. ~5,000 registered users and ~CHF 50,000 in revenue in 2024.

Motor anomaly detection.

Detecting rare defects in precision motors from noisy current, voltage and temperature signals. 92% recall, on the production line.

Cancer organoid detection & tracking.

Tracking organoids to decide when they're ready for drug testing, with a human-in-the-loop pipeline that automated 89% of annotation.