projects

Hand-written CUDA kernels for LLM inference on an RTX 5090: RMSNorm, SwiGLU, softmax, fp32 and bf16 GEMM and fused attention, each as a ladder of variants benchmarked against cuBLAS and PyTorch and profiled with Nsight Compute. The bf16 GEMM runs ahead of cuBLAS from 2048 cubed up with TMA and Stream-K rungs, the decode GEMM at the memory floor, and attention ahead of FlashAttention-2.

AI assistant for small businesses, built with Silkroad Innovation Hub. Every chat gets triaged and distilled into a wikilinked Markdown vault behind an approval queue; reaches owners over iMessage and email with a budget monitor, audit log, and trust ramp in front of anything that sends.

Swarthmore's self-hosted campus LLM, run by SCCS. Qwen3.6-35B-A3B on vLLM, hybrid dense and sparse retrieval over 3,500 scraped campus docs with a 1.5 second fail-open budget, live campus data over MCP, Keycloak login. Nothing leaves campus.

Parameter recovery for a 7-state colorectal cancer ODE model with physics-informed neural networks: an integral residual that beat autodiff 50 to 37, Fisher information and Bayesian PINNs for identifiability, and neural-mechanistic hybrids that only learn the mechanism when an experiment drives its input to zero. Paper with Nate Kim and our advisor.

Voice-driven AR coaching over a live camera feed, top 6 at xAI's Grokathon. Web Speech in, Grok 4.5 for vision and answers, xAI TTS out, three hedged requests to make box calls fast, and a dependency-free TypeScript tracker to keep them on the object.

Task and project management for student orgs, on SCCS infra. Next.js server actions, Prisma on Postgres, a Swarm service pinned to one node, and no auth yet.

Social platform built around intentional sharing. Five items a day, 24-hour expiry, group-based feeds, designed against engagement maximization. Half-built on Supabase.