Projects

Log of built work

Projects, newest first
Year	Project	Status	Built with
2026	Bable: Research RAGA research assistant that discovers, retrieves, compares and synthesises scientific literature from multiple academic sources, with hybrid retrieval and answers grounded in citations.	Status: in progress	Built with: Python, Retrieval-augmented generation
2026	CUDA GEMM JourneyA from-scratch CUDA single-precision matrix multiplication (SGEMM), optimised step by step and benchmarked against cuBLAS, PyTorch and CUTLASS. Each kernel generation starts from a profile in Nsight Compute, fixes the one bottleneck it finds, and is backed by measured timings, GFLOPS and hardware counters rather than estimates.	Status: in progress	Built with: CUDA, C++, Nsight Compute, cuBLAS, CUTLASS
2026	CUDA Vector AdditionAdds two vectors of 100 million elements on the GPU and times the kernel: a first CUDA program covering global index calculation, thread and block configuration, and host-device memory transfer.	Status: complete	Built with: CUDA, C
2026	GPU Electron Velocity SimulationSimulates electron velocities on the GPU with NVIDIA Thrust: millions of normally distributed samples are reduced to a mean velocity, showing it approach zero as the particle count grows, a small reproducible demonstration of statistical equilibrium.	Status: complete	Built with: CUDA, C++, Thrust
2026	GPU Neighbor Gather: Divergence vs. CoalescingA benchmark of the neighbour-gather pattern at the heart of particle simulations and graph neural networks. It varies how neighbour counts are distributed across elements and measures whether warp divergence or the memory-access pattern dominates the cost, a question raised by OpenFPM's claims of near hand-tuned SPH throughput.	Status: complete	Built with: CUDA, C++
2026	NodeGuardA graph neural network that flags coordinated mule-account rings by looking at who an account is connected to, not just its own numbers. Built in a few days for Razorpay's AI Buildathon: recall 0.333 and precision 0.200 on a held-out test set with only three mule accounts, so evidence that the approach is real rather than a production benchmark. The demo is a static site built from the trained model's real predictions.	Status: complete	Built with: Python, Graph neural networks
2026	PagedServeA from-scratch, high-throughput LLM inference engine with continuous batching and a paged KV cache.	Status: in progress	Built with: LLM inference, Paged KV cache, Continuous batching
2026	RISC-V Knowledge DBA small, queryable database built from the RISC-V Unified DB: its YAML specifications become JSON snapshots and then PostgreSQL tables, so questions like which extension defines an instruction, what an opcode's encoding is, or which CSRs exist at which privilege level become simple queries.	Status: maintained	Built with: Python, PostgreSQL, RISC-V
2025	Backprop in PythonA hands-on reimplementation of micrograd, built to learn how neural networks and backpropagation work from scratch.	Status: complete	Built with: Python, Jupyter

Practice, dailyChecked 26 Sept, 19:51 UTC

GitHub262 contributions in the last year

LeetCode573 submissions in the last year

Codeforces40 submissions in the last year

Résumés

GPU Programming · HPCCUDA kernels, Nsight profiling, memory coalescing and shared-memory tilingDownload the GPU Programming · HPC résumé
AI / Machine LearningPyTorch, graph neural networks, retrieval-augmented generation, agentic pipelinesDownload the AI / Machine Learning résumé
Research · AcademiaMachine learning for materials science, language models and high-performance computingDownload the Research · Academia résumé