Projects

LLM customer with a hidden goal, 13-tool agent, reward gated on the database diff before a judge checks consent and honesty. Red-teaming caught a forged-consent hole in the judge.

Induced sycophancy in Qwen3-8B, then compared DPO, SimPO, IPO, GRPO and Constitutional AI. DPO fooled the judge while a probe still found the disposition; CAI preference pairs did best.

Co-founder and CTO. Multi-tenant SaaS that tracks competitors’ launches, discontinuations and price changes, with LLM output checked against live search.

Evaluator for an account-risk agent: 29 spec rules checked against corrected telemetry and source text. 259 of 629 numeric claims failed to reproduce.

Production legal document processing system using LLM-as-a-Judge framework, achieving 87.5% accuracy with 60% time reduction.

Custom PDF parser with fine-tuned LLMs for aerospace technical documentation, achieving 80% effort reduction and securing $200k funding.

Research using LLMs to evolve agent behaviors through automated code generation. Accepted at GECCO ‘25.

Extension of LEAR research exploring quality-diversity tradeoffs in agent evolution.

Extension bridging NetLogo and LLMs for intelligent agent behaviors and decision-making.

Claude Code plugin for the job hunt: tailored LaTeX resumes, live ATS form schemas, and a judge that checks every claim against your own profile file.