Product work, experiments, and research with focused notes and visuals.
Profiling MLA attention on H100 reveals reconstruction GEMMs consume 61% of attention-layer time. INT4 quantization should help but doesn't, because the weights fit in L2 cache.
Packed INT4 GEMM kernel in Triton for decode-heavy LLM inference with hardware counter attribution and a regime model for when quantization helps. Up to 3.7× speedup on A10G.
Kernel development and HW/SW co-design for Atalla, a student-built weight-stationary systolic array AI accelerator. FlashAttention mapping, im2col convolution, tiled GEMM, and PyTorch backend integration.
Fully pipelined 5-stage RISC-V processor in SystemVerilog with hazard detection, data forwarding, and a 2-bit branch predictor. Synthesized in Cadence targeting 70 MHz.
A scheduler for ML accelerator computation DAGs that minimizes latency while respecting tight on-chip SRAM capacity constraints. Google MLSys 2026, Track A.
A fused CUDA kernel implementing DeepSeek Sparse Attention on NVIDIA B200 (Blackwell). FlashInfer AI Kernel Generation Contest, MLSys 2026.
Integrated RTL design verification tool for testbench generation, script and trace analysis with multi-agent collaboration for accelerated RTL development.