Long-context LLM systems
KV-cache compression, sparse attention, distributed attention, and evaluation pipelines for scaling context length under real compute constraints.
Research Focus
I work on making long-context and diffusion language models more efficient without giving up the behavior that makes them useful.
Masking and feedback dynamics
Studying how remasking, interpolation, and embedding geometry affect convergence, hallucination lock-in, and fine-tuning stability.
Vision and multimodal systems
Computer vision, synthetic EO data, thermal imagery, retrieval systems, and civic intelligence dashboards.
Current Work
Research, systems, and applied ML that ship cleanly
I care about models that are not only interesting on paper, but also efficient, inspectable, and usable in real workflows.
Efficient inference
KV-cache compression, sparse attention, distributed long-context systems
Diffusion LLMs
Masking dynamics, interpolation behavior, and decoding stability
Applied ML builds
Vision pipelines, research agents, civic dashboards, and deployment-ready tools
Featured Publications
A small selection of recent work. The full publication list has links and BibTeX entries.
More Than a Quick Glance: Overcoming the Greedy Bias in KV-Cache Compression
- Introduced LASER-KV, a KV-cache compression method improving long-context accuracy over greedy pruning baselines via block-wise accumulation and Exact-LSH sampling.
Dropping the Anchor: Statistical Summarization for Distributed Systems via Pulsar Attention
- Designed Pulsar Attention for distributed long-context inference with content-aware statistical summaries and lower Phase-1 FLOPs than Star Attention.
Experience
Research and engineering roles spanning long-context LLMs, distributed inference, and computer vision.
Shantanu Acharya, Research Scientist at NVIDIA
Research Intern- Designed and implemented Pulsar Attention, a distributed long-context LLM inference framework that replaces static anchor duplication with content-aware statistical context summaries.
- Built the experimental stack for distributed attention, Max-IDF summarization, FlashAttention-based inference, and evaluation on RULER and BABILong.
- Co-authored a paper showing up to 4.7% accuracy improvement over dense attention, 6.1% over Star Attention, and 3.3x lower Phase-1 FLOPs.
Auric AI Labs
Computer Vision Intern- Created an end-to-end synthetic EO data generation pipeline for object-detection models using controlled top-view renders from 3D military-object assets.
- Developed a GPT-assisted semantic placement pipeline using probabilistic spatial priors to identify contextually valid object locations.
- Combined BlenderProc rendering, Meshy AI asset generation, and RemoteCLIP-based semantic filtering to improve synthetic scene realism.
Selected Projects
Highlights from the larger project archive, now moved to a dedicated projects page for readability.
ParkWatch
AI-powered Bengaluru parking-enforcement intelligence dashboard with GraphSAGE hotspot forecasting, Mappls road-aware A* patrol planning, delay-exposure estimates, and exportable action plans.
Research Corpus Agent
Multi-agent scientific-literature RAG over 100K+ arXiv papers with hybrid retrieval, reranking, multimodal figure/table understanding, and evaluation pipelines.
GATE-IR
Gated thermal perception pipeline routing infrared images through weather-specific enhancement before lightweight YOLO-based detection.
Education
Indian Institute of Technology Roorkee
Bachelor of Technology in Electrical Engineering- * US scale conversion calculated via Scholaro methodology.
Recent Highlights
Selected milestones across research, hackathons, and campus activities.
- 2nd Place, PixelPlay 2025 – campuswide computer vision hackathon.
- Project selected for IIT Roorkee Institute Research Day 2025 through the TMI program.
- Top 400 of 10,000 teams and selected for round 2 of Flipkart Gridlock 2.0.