Bennett Lee

12.7% domain gain · 10,000+ personnel · ~20% workflow improvement

Fine-tuned a large language model for legal work at Singapore's Home Team Science and Technology Agency. The models went into AI tooling deployed to over 10,000 public-safety personnel, improving tested workflows by around 20%.

Investigated catastrophic forgetting across LoRA, CPT, and full-parameter SFT, lifting legal-domain performance 12.7% while holding general-capability loss under 3%. Trained across 8× H100s with FSDP2, sequence packing, and fused kernels.

20+ pledges · 200+ members

Helped develop the pledge-class overview for Kappa Theta Pi's app, Michigan's premier technology fraternity: a live points ring, leaderboard, activity log, and strikes tracker. The app is used by rushees and over 200 members for everything to do with the fraternity.

Designed the UI in Figma, built it in React and TypeScript, and migrated the feature off hardcoded demo data onto Firestore so class progress syncs across users in real time.

A multi-agent system that takes a complex request, breaks it into subtasks, and routes each one to the right specialist agent, research, web search, email, or sandboxed code execution.

LangGraph orchestration with SQL task memory, semantic caching over embeddings, Arize Phoenix tracing, and network-isolated ephemeral Docker containers for running model-generated code safely.

A harness that makes LLM agents trustworthy, reliable, and useful for the specific jobs you point them at.

Framework-free, with schema-validated tools, permission gating and human approval, persistent tracing, deterministic replay, failure injection, and automated regression evaluation across runs.

A free tool helping Singaporean families find the eldercare subsidies they qualify for. Nearly half of caregivers report not knowing what support exists, while the schemes themselves sit behind a maze of overlapping eligibility rules.

A deterministic rules engine over a hand-curated dataset of official schemes. Chose plain logic over an LLM so eligibility guidance stays accurate, explainable, and traceable to the official source.