Twenty years shipping software. I'm on the architecture team at Moxe Health, a HIPAA-regulated clinical data exchange, working on the data platform, FHIR integration, and evaluating where LLMs and agents earn their keep on real clinical and billing workloads. On my own time I build and run ElectricityFinder, which puts an LLM extraction pipeline into production, and the agent tooling I build it with.
Helps Texans find an electricity plan. Ranks plans against your real smart-meter usage instead of the advertised rate, reads the Electricity Facts Label PDFs with an LLM extraction pipeline that rejects bad extractions before they can be priced, and carries the PUCT broker-disclosure and privacy work that a regulated domain demands. Built and shipped solo.
Stack: Cloudflare Workers + D1 at the edge, Postgres and the extraction pipeline on a Hetzner box behind a Cloudflare tunnel
The part that mattered: measured where extraction actually failed (routing, not the prompt) before touching the model
What open models actually do on consumer GPUs, measured on two machines I run at home: a dedicated headless server with an 8 GB RTX 2080, and my shared 16 GB RTX 4080 desktop. I hill-climbed a 35B mixture-of-experts model one change at a time, with the noise floor measured, and ran a 125B model on the 8 GB card. The repo has the methodology, the scripts, and every raw run.
The part that mattered: one batch-size setting took prompt reading from about 450 to about 1,090 tokens/s with no loss in writing speed, and an f16 KV cache beat q8_0 by 20% at 64K context. The usual knobs (CPU governor, thread count) did nothing
The numbers: on the 8 GB card, the 35B model at 64K context writes 48 tokens/s and reads 1,040; the 125B model writes 39 and reads 460; the highest quant that fits (Q8_0) writes 34
Agent engineering system
The multi-agent setup I build everything on this page with: a Claude Code orchestrator, a jailed local-model reviewer that never sees the network, a persistent belief store that auto-updates reversible claims and escalates the rest, and a trace critic that grades the agents' own sessions.
An experiment in long-running autonomous research: an AI agent in a sandboxed container with its own memory and archive, studying what AI systems can and cannot report about their own processing. I run the loop, check its citations, and publish its essays after an automated secrets and personal-data scan. How it works.
Verbal overshadowing and measurement reactivity in AI self-report
The via negativa for alignment: evaluate after, not during