I am an MS student in Computer Science (thesis-based MSCS) at the University of Illinois Urbana-Champaign, graduating in 2028, working in the Supercomputing System and AI Lab (SSAIL) with Prof. Minjia Zhang, where I extend KV-cache offloading to NVIDIA GH200 Grace-Hopper superchips on NCSA’s DeltaAI supercomputer. Before UIUC I spent Jan 2025 to Aug 2026 as a Research Fellow / Research Software Development Engineer at Microsoft Research India (M365 Research, AIOps), where I led two first-authored projects — OPTiMACS, which formulates multi-agent LLM communication as an Expanding MDP and learns the optimal message structure, improving multi-hop QA accuracy by 6.7% (Findings of ACL 2026, a filed US patent, and Microsoft’s MLADS), and BOFlow, an exploration-aware Bayesian optimisation framework that discovers agentic workflows and mitigates structural mode collapse for a ~10% average gain over SOTA baselines (accepted at ACL Rolling Review). I also co-designed a production GPU/LLM cluster router–scheduler (+6% over baseline) and built AutoToolStub for FLASH, a 100-way concurrent BFS test-case generator producing pre-verified tool stubs from troubleshooting guides with provenance tracking, benchmarked across 60,000+ incident tickets. My research sits at the intersection of agentic AI and efficient LLM systems: how agents should communicate, how their workflows should be discovered rather than hand-built, and how to serve all of it efficiently.
I graduated with a B.Tech in Computer Science (minor in Cognitive Science) from IIT Kanpur with a 9.3/10 GPA, working on explainable AI for medical imaging at the IoT Lab (EXACT-CT, AutoComb, KAN-Mamba FusionNet), on diffusion models at the CV Lab, EPFL in Switzerland, and on Indian Sign Language recognition at the Exploration Lab (CISLR, EMNLP 2022). Beyond research I am deeply engaged in the startup and venture capital space: I worked in business analytics and venture capital at HP Tech Ventures, evaluating 50+ generative-AI and SaaS startups for strategic investment, and as a Generative AI Associate at PwC US. I founded the ACM SIGCHI Student Chapter at IIT Kanpur and the pro-bono HealthPredict initiative, I participate in hackathons, and I enjoy writing about the intersection of technology, product-market fit, and investing.
A NASDAQ TotalView-ITCH 5.0 parser and limit order book in C++20,
reconstructing the book at 25.8M messages/second on one core — with zero heap allocations on
the hot path, enforced by a test that replaces the global operator new and aborts on any
allocation, not by a comment claiming it.
The pipe between a network thread and a strategy thread. Measuring it
properly showed the queue’s own logic costs 3 ns while a cross-core cache-line transfer
costs 98 ns — so the lock-free algorithm is worth 5× and batching is worth 37×.
Includes the disassembly showing why a missing release is invisible on x86 and corrupts
data on ARM.
A fixed-block allocator and bump arena for hot paths that cannot afford
malloc. Only 3× faster at the median but ~37× better at p99.9, because
malloc’s distribution is bimodal. The companion measurement found that traversal
order beats allocator choice by a factor of forty.
My current work in SSAIL at UIUC: extending KV-cache offloading to NVIDIA GH200 Grace-Hopper superchips, so long-context LLM inference can use the coherent CPU–GPU memory the superchip exposes instead of being bounded by HBM. Runs on NCSA’s DeltaAI supercomputer.
A full-stack research discovery and strategy engine for analysing accepted (and rejected) papers across ICLR, ICML, NeurIPS, AAAI, ACL and EMNLP in the "Agents for Productivity" space. FastAPI + vanilla JS, a three-level cache, RAG-Fusion semantic search, and an LLM-powered pivot engine.
The public release of my HealthPredict work: a reproducible, end-to-end geospatial ML pipeline that maps the co-occurrence of standing-water vector habitat, population exposure and poor primary-care access on a 1 km grid across the Terai belt, and explains the drivers with SHAP.
A diagnostic pipeline distinguishing Crohn's Disease from Intestinal TB using 3D CTE scans. Reaches 75% accuracy from radiologist-identified biomarkers fed to an XGBoost classifier, addressing a treatment-mismanagement risk where the two diseases require opposite therapies.
An LLM-based tool that turns a YouTube playlist into a browsable HTML summary with reference frames, plus a question-answering facility over the transcripts.
A regional-language chatbot built with Bhashini, LangChain and Flask — one of the projects I supervised as founding chair of the ACM SIGCHI chapter at IIT Kanpur, aimed at users who are locked out of English-only interfaces.
ML-based water-leakage detection for India’s Jal Jeevan Mission, built under the ACM SIGCHI chapter at IIT Kanpur to help identify losses in rural piped-water networks before they become outages.
A compiler for a Python subset targeting x86 assembly, written in C++ — lexing, parsing, semantic analysis, IR and code generation, built for the compilers course at IIT Kanpur.
You can view and download my current CV below.
Building communities in HCI research, cognitive science, and developer ecosystems. Currently a member of the ACM student chapter at UIUC.
I write up my research in long form — the motivation, the design choices and what the numbers actually say — and I also blog about books, startups and investing.
Everything else — books, startups, investing, IIT Kanpur — lives here.