Projects & systems

I run my own large language model on a single desktop machine — no cloud, no external API. I set up the model, the serving, and the tooling so it can be driven like any other API. On top of that I built an automated pipeline that turns 80 work records into a deduplicated, evidence-traced memory where every output carries its source. The hard part isn't the model — it's making a model that can be wrong produce output you can trust and trace. Under the hood: an NVFP4-quantized Qwen 27B model, a 262K-token context window, and an OpenAI-compatible API via SGLang (migrated from vLLM behind the same contract).

• A 27B model running locally on one 128 GB desktop box — no cloud
• 262K-token context, exposed as an OpenAI-compatible API
• Automated pipeline: 80 records, zero missed after fallback, every output traceable to its source
• Hybrid deduplication that only escalates ambiguous cases to the model

This app classifies email across all your connected accounts (iCloud, Gmail, Yahoo, work) entirely on your Mac — nothing is sent to a cloud. You define folders in plain English, and it files mail accordingly, learning from how you actually sort things over time. A companion iPhone app lets you approve or reject decisions and watch it run live. It's a product where the model is just one component — the privacy architecture and the decision logic are the real engineering. Under the hood: Swift, Apple's on-device Foundation Models by default (or open-weight models you run locally), with multi-account support.

• Zero email content leaves the machine — local models only
• 30,000+ of my own emails classified across multiple accounts
• Folders defined in plain English, and it learns from your corrections
• Companion iPhone app: approve/reject and live status, with no on-phone inference