Independent projects

Things I built
to find out.

Projects I started myself, usually to find out whether something was possible. A home lab, a self-hosted finance ledger, a GPU sharing platform.

Self-hosted GPU sharing platform March 2026

GPUShare

Built a self-hosted platform for trusted users to share local AI inference and Blender rendering, with optional cloud routing and estimated usage-cost allocation.

Result The application implements auth, local and optional cloud inference, MCP tool routing, queued rendering, and append-only usage accounting. Host availability and a stale installer migration command remain operational limits; no adoption claim is made.

View case study Source on GitHub
Open-source finance ledger 2026

You Inc

Built a personal-finance ERP with a double-entry ledger, Akahu bank sync and configurable dashboards, then took the commercial layer back out and made it self-hosted and MIT licensed.

Result A working product you install rather than sign up for. youinc.net now serves a sample-data demo and the documentation; the software itself is on GitHub under MIT, and anyone running it points it at their own database and their own Akahu credentials.

View case study Source on GitHub
Apple Health data pipeline January 2026

HealthAgent

Built a pipeline that normalises daily Apple Health exports into Postgres and produces personal trend views.

Result A working end-to-end ingestion and trend pipeline; generated insight backends remain optional and are not medical advice.

View case study Source on GitHub
Raspberry Pi recovery architecture 2025

Home Lab

Designed and tested a resource-constrained home-lab stack with private administration, public access, and encrypted offsite recovery.

Result A full restore completed in 1 hour 45 minutes; 96GB of backup storage cost $2.82 per month at the recorded point in time.

View case study
Academic machine-learning study June 2025

Wildfire PySpark

Tested a distributed Spark ML pipeline on 1.88 million United States wildfire records.

Result Class weighting raised rarest-class recall from 0 to 0.7518, but overall accuracy fell to 32% and rarest-class precision to 0.0099.

View case study