AI/ML engineer and applied researcher working on evaluation infrastructure, agent frameworks, and RAG systems. I ship open-source used in production by teams at MongoDB and LINE. I also work on AI safety, mostly hands-on: fine-tuning model organisms with hidden objectives and building detectors for them. MCS student at UIUC.

• inspect_ai: the UK AI Security Institute runs its frontier model evaluations on this framework. I work on its agent bridge, the piece that lets outside coding agents be scored against Inspect models. My work there spans all three provider integrations: OpenAI, Anthropic and Google.
• trustgraph: production RAG and agent framework. I built out its retrieval and multimodal layer: cross-encoder reranking, hybrid BM25 and vector search, source attribution, and vision input. All of it shipped in production releases.
• comprisk: competing-risks survival analysis for Python, and my first-author research (arXiv:2607.09431). The only native-Python competing-risks random survival forest, bit-identical to R's randomForestSRC and 10-22x faster.
• js-toml: TOML parser and serializer for JavaScript. 18k+ weekly downloads with 3.8x year-over-year growth, TOML 1.1, passes the official toml-test compliance suite. Used by 400+ projects, including MongoDB's docs platform and LINE's user feedback product.
• omnigent: Databricks' open-source agent framework and meta-harness for orchestrating Claude Code, Codex and Cursor. I contribute to its runner and web UI.
• Model organisms and detection: I fine-tune models to serve a hidden principal, then try to catch them. Blind weight-space remediation removes the loyalty about a quarter of the time and destroys the evidence of who it served nearly every time. A detector that scored near-perfect turned out to be reading how the models were built rather than the backdoor itself. Code and numbers in secret-loyalties-testbed.
• AgentQuant: open-source AI agent platform that turns natural-language trading ideas into quantitative backtests. Top 3 at the NVIDIA AI Hackathon, 2025. Built at NLPQuant, the AI studio I founded.
• Led Mogara, the organization behind QSanguosha, an open-source implementation of the Sanguosha board game with millions of downloads; its Lua API supported a community of 1000+ mod creators.