projects

Amalgamation of various short and long-term undertakings.

some machine learning

  • [Re]FlashAttention - Triton Implementation of FlashAttention1 with Multi-Query Attention (MQA). [White Paper] [Poster]
  • LLM Homogenization in Creative Writing - Investigated claims of LLM homogenization in creative writing; our work shows that with sufficient initial context, LLM-generated text corpora achieve stylistic and semantic diversity on par with human writers. [White Paper]
  • Astralis - Contributed to an early-stage startup using AI and geospatial data to optimize real estate development and demand forecasting. Built intelligent document ingestion pipelines (RAG) using Python, LLMs, and ML clustering with automated metrics and async workflows for scalable content organization.

more kernels and numerics

  • [Re]Spectral Methods - An implementation of "A Fast and Well-Conditioned Spectral Method" (Olver and Townsend, 2013). [White Paper]
  • GPU MODE! - wrote some kernels in Triton for GPU MODE competitions
  • Butterfly Matrix Kernel – Optimized implementation of butterfly matrix multiplication using Triton, enabling fast structured transforms for deep learning workloads.

hacks

  • Propagation of In-Context Scheming in Multi-Agent Systems - Designed experiments to measure how corrupted instructions spread through networks of language models, demonstrating vulnerability of small models to influence from both peers and higher-capability models. Awarded 2nd place at Anthropic Alignment Hackathon (2025).
  • SearchDestroy - Robust multi-drone search algorithm (DARP + heuristics) for area sweeping, handling drone loss and online path re-computation. Physics simulation & visualization using AirSim. Awarded 2nd place (100+ participants) at AGI House Robotics Hackathon (2024).

bio/chem/neuro