Projects

Research systems, experiments, and tools. Each entry is kept as a working note.

auto-kernelgpu-kernels · memory-bandwidth · llm-trainingAutonomous roofline-guided GPU kernel optimization. Profiles whether a kernel is compute- or memory-bound, proposes a hypothesis from the optimization guide, modifies the kernel, benchmarks against the reference, and iterates — committing, benchmarking, and reverting autonomously.Source
beaconllm-training · retrieval · benchmarks · applied-agentsContext assembly for LLM code repair under strict token budgets. Parses Python repos into a knowledge graph (AST → calls, imports, inheritance), runs BM25/dense/BCA retrieval, assembles context that fits the budget, and measures whether better retrieval actually fixes more bugs.Source
tierra+bayesian-inference · scientific-computing · astrophysics · retrievalCalibration-first retrieval stack for exoplanet atmospheres + NEOs. Forward-model wrappers, opacity ensembles, contamination latents, and reproducible inference in one package.Source
gymforgereinforcement-learning · benchmarks · applied-agents · evaluationReproducible, evaluable reinforcement learning environments for training agents on computer-use tasks. Structured framework for generating diverse task batches, running sandboxed episodes, and scoring agent performance.Source
compressmodel-compression · quantization · llm-training · efficiencyActivation-aware compression for transformer models. Profiles which attention heads and MLP neurons are most active on your data distribution, then applies more aggressive compression to dormant ones — better task-specific quality at the same compression ratio.Source
The Third Pillarreinforcement-learning · applied-agents · social-dilemmas · benchmarks · multi-agent-learningMulti-agent simulation measuring LLM strategic behavior in resource competition under asymmetric information and institutional constraints.Source
pufferXvector-search · databases · ivf-indexing · object-storageObject-storage-first hybrid search engine w/ IVF vector search + BM25 full‑text with metadata filters, immediate write visibility via a hot index, and async merge to an S3‑backed cold index. Features adaptive nprobe, NVMe caching, and RRF fusion for robust relevance.Source
inductramechanistic-interpretability · in-context-learning · transformer-architecture · attention-mechanismsMech interp playground for tracking ICL emergence and induction heads. Built to understand what transformers actually learn during in-context learning.Source
shardsparkdistributed-training · parallelism · fsdp · gpu-kernelsMinimal FSDP implementation. Stripped down to the essentials so you can actually see how parameter sharding works under the hood.Source
GPU-Accelerated-Denoisergpu-kernels · memory-bandwidthCUDA-accelerated BM3D denoising. Gets you ~20% speedup over OpenCV and handles real-time video.Source
kaironvector-search · databases · hnsw-indexingVector db built from scratch. HNSW, KD-tree, IVF- multiple indexing strategies with proper metadata filtering because most vector DBs overcomplicate the basics.Source