AI engineer building systems that separate what should be automated from what is uniquely human, empowering people to use AI and focus on what only they can do. BS/MS student in Computer Science (Artificial Intelligence concentration) at Stanford.
| Project | What it asks | Stack |
|---|---|---|
| emergent-cooperation-pandemic | Can selfish multi-agent RL agents learn to share scarce medical resources during a pandemic? An impact reward cuts deaths 4–11% vs. selfish baselines. · CS 224R | PyTorch · IPPO / MAPPO · SEIR sim |
| cot-faithfulness-audit | Is chain-of-thought causal, or post-hoc rationalization? Logit lens + activation patching on Llama-3-8B. · CS 281 | TransformerLens · Llama-3 |
| showjumping-ssl | Predicting showjumping outcomes from broadcast video with self-supervised video representations, controlling for the venue confound. · CS 131 | R(2+1)D · YOLOv8 · InfoNCE |
| mcphases-rl | Offline RL (FQI + behavior cloning) for personalized menstrual-health recommendations from wearable data. · CS 238 | scikit-learn · offline RL · WIS |
| HeuristicWithQLearning | Can Q-learning learn better A* heuristics for imperfect mazes with loops? · CS 221 | Python · Q-learning |
- Civitas: helps small government contractors find compatible RFPs across 62 California procurement sites using semantic matching. Led a 5-person team, ran 10+ user discovery interviews, and wrote most of 640+ commits. Next.js · Postgres + pgvector · AWS Lambda · Playwright
- Rooted: an AI decision partner and decision history for startup product managers. It pulls context from where the team already works, helps weigh the options, and keeps a searchable record of what was decided and why. Shaped by 16 interviews with product managers. In progress with Team Geckos for Stanford CS 177. Landing page. Vite · React · Supabase · Vercel
- SAIPIENS: a novel AI homework system that grades students' reasoning, not just their answers. Teachers upload curriculum PDFs, the system extracts learning objectives, and it generates units for students, using research-backed pedagogy to cut teacher and student workload. Top contributor (155 commits). Stanford CS 224G. Live app. React · Vite · Tailwind · AWS Lambda + Cognito · Gemini
- Prose.ai: a website builder where plain-English specs compile into tested code. In progress, private repo.
- BookClubB: an iOS app for book clubs with groups, discussion threads, and profiles. SwiftUI · Firebase
- ReelLife: a narrative Unity game with day-by-day schedules, character stats, branching dialogue, and multiple endings. Built with Team Reithrodon for Stanford CS 247G. Unity · C#
Two merged PRs to crewAI, the multi-agent orchestration framework: an offline evaluation runner for agent routing and guardrail decisions (#8230, 46 tests, mutation-tested) and credential-free AWS Bedrock contract tests (#9096). I also filed #9095, a Bedrock Converse bug.
Tools: Python · PyTorch · TypeScript · Swift · SQL · AWS
How I build: Claude Code, Cursor, Factory; my own agent skills and MCP servers. Every public repo has an AGENTS.md, one-step setup, and tests an agent can run.