Fused Triton kernels for ColBERT MaxSim with dimension tiling (d>128) and fused PQ scoring. Reached 80% peak HBM bandwidth on H100 via roofline-guided tiling, 1.9× over PLAID's GPU kernel, 469× exact-scoring throughput over WARP (SIGIR'25 Best Paper) and 8.5× over torch.compile.
Triton
CUDA
GPU Kernels
Information Retrieval
GPU-native inverted index and fused scatter-add kernel scaling exact SPLADE retrieval to 8.8M docs. 235× speedup over Pyserini CPU and 23–270× over SPARe's GPU scatter-add kernel. Characterized the work- vs. bandwidth-efficiency tradeoff governing GPU sparse retrieval kernel design.
Triton
CUDA
Sparse Retrieval
GPU
Designed a metric for quantifying ordering anomalies in batch-ordered logs. Implemented Windowed Temporal Reordering using Lamport timestamps, reducing causal anomalies by 80% while preserving throughput.
Go
Distributed Systems
Lamport Timestamps
Implemented a decentralized total ordering algorithm for transaction processing with fault-tolerance via heartbeat-based failure detection and reliable multicast.
Go
Distributed Systems
Fault Tolerance
Developed a Rate Monotonic Scheduler for real-time task scheduling. Optimized memory allocation with a Slab Allocator and implemented a kernel module for profiling page fault rates.
C
Linux Kernel
Systems Programming
Modified attention mechanisms using dependency parsing to encode syntax. Deployed a dual-encoder neural retriever with Elasticsearch, achieving 44.5% MAP@10 and 49.7% MRR@10 gains over ColBERTv2.
Transformers
Elasticsearch
NLP
Implemented a custom TCP protocol over UDP with ACK-based retransmission and AIMD congestion control including Slow Start, Congestion Avoidance, and Fast Recovery.
C++
Networking
Congestion Control
Developed a music analyzer extracting nearest frequencies using Fast Fourier Transform. Parallelized with OpenMP and CUDA for CPU and GPU acceleration.
C++
CUDA
OpenMP
