Agent evaluation and harnesses

I build systems for supervising long-running agents: bounded execution environments, scoped tool access, reversible writes, verification checks, approval gates, and readable transcripts for debugging failures.

My core engineering principle is simple: delegate tasks that are inexpensive to verify. A difficult task with a reliable automated check is often safer to delegate than a simple task requiring complete human review.

Research compute infrastructure

I built and operate a 4U GPU server. The system supports local inference for Llama-class 70B models, LoRA and QLoRA fine-tuning, embeddings, reranking, vision workloads, and concurrent smaller-model serving.

Wrodium's automation platform

As co-founder and CTO of Wrodium, I built marketer-facing workflows that could crawl a website, evaluate its visibility in AI answer engines, identify structural problems, generate changes, route them through human approval, publish through six CMS integrations, and monitor subsequent citation performance.

The central architectural challenge was translating inconsistent CMS content into a common representation without losing meaning. We moved toward maintaining a verified source of truth about each business and generating content from that structured knowledge.