I research how to make long-context language models cheaper to run — KV-cache compression, sparse and distributed attention, and the dynamics of diffusion LLMs.
B.Tech EE ’28 · IIT Roorkee
( 01 — About )
Electrical Engineering undergraduate at IIT Roorkee, working where systems meet models. I build methods that let language models read longer contexts for less compute — and I care that they stay inspectable, reproducible, and usable in real workflows.
( 02 — News )
Are You Sure You're Sure? accepted as an oral at the JUDGe Workshop, NeurIPS 2026.
BRIEF and Underscoring the Problem accepted at the Long-Context Foundation Models (LCFM) Workshop, NeurIPS 2026.
Joined Simplismart as an ML Engineer Intern, working on pipeline- and kernel-level inference optimisation.
Dropping the Anchor (Pulsar Attention) accepted to Findings of EMNLP 2026.
Joined Lexsi Labs as an AI Research Intern, working on long-document generation (BRIEF).
Lost in Interpolation accepted at the Non-Autoregressive LM Workshop, COLM 2026.
Started an independent research internship with Shantanu Acharya (Research Scientist, NVIDIA) on distributed long-context inference.
More Than a Quick Glance accepted at the ICBINB Workshop, ICLR 2026.
Don't Look Up (Every Token) published in the ICLR 2026 Blogpost Track.
Preprint of LASER-KV released on arXiv.
( 03 — Research )
★ Highlight · EMNLP 2026 Findings
Star Attention prepends a full copy of block 1 to every GPU; Pulsar sends a 64-token sink plus Max-IDF chunks (rare tokens) from earlier blocks, then discards that prefix and merges the query globally.
Interactive explainer
( 04 — Experience )
Simplismart
Working on LLM inference optimisation for production serving, with research spanning both the pipeline level and the kernel level.
Lexsi Labs
Worked on long-document generation and led the research project BRIEF (Bounded Regeneration In Exact Form), which regenerates long legal and financial documents clause by clause from short structural briefs derived from a reference graph.
Shantanu Acharya, Research Scientist at NVIDIA
Designed and implemented Pulsar Attention, a distributed long-context LLM inference framework that replaces static anchor duplication with content-aware statistical context summaries.
Auric AI Labs
Created an end-to-end synthetic EO data generation pipeline for object-detection models using controlled top-view renders from 3D military-object assets.
( 05 — Projects )
Source-grounded long-context fact-QA harness with length and position sweeps, distractor and multi-hop cases, exact-span hallucination checks, and calibration reports.
Parking-enforcement intelligence for Bengaluru traffic teams: GraphSAGE hotspot forecasting, patrol sequencing, and exportable action plans.
Multi-agent literature RAG with planner, retriever, analyst and critic agents, hybrid retrieval, and figure/table understanding.
Early-fusion Transformer over 112-dim multi-channel GNSS sequences with physics-informed features.
Gated thermal perception: weather-specific enhancement routed before lightweight YOLO detection, at near-YOLO latency.
Diffusion-LLM decoding with confidence momentum, semantic hysteresis and entropy-aware control against hallucination lock-in.
A lightweight neural token pruner trained on attention distributions that drops 60%+ of context before inference.
( 06 — Writing )
( 07 — Education & Recognition )
2024 — 2028
Bachelor of Technology, Electrical Engineering
8.175/ 10 CGPA · 3.538 / 4*
* US-scale conversion via Scholaro methodology.
2nd Place, PixelPlay — campus-wide computer vision hackathon
Institute Research Day — project selected via the Tinkering & Mentoring program
Flipkart Gridlock 2.0 — top 400 of 10,000 teams, round 2
JEE Advanced AIR 1519 · JEE Main AIR 2912