Anupam Rawat CV

Research engineer in Mumbai,

I make language models decode faster.

1.35× faster decoding at Conscious Engines. Before that, AI for physics at IBM Research.

Open to full-time research engineer roles from mid-2027.

Anupam Rawat
Paper · NeurIPS 2026

COSMOS

47%

lower long-horizon rollout error than baseline neural operators. Built at IBM Research.

Read on arXiv ↗
Product · 2026

Whettle

Live

at real hotels. I was the only engineer: bookings, billing, payments and guest messaging.

Hackathon · Sep 2026

Amazon ML Challenge

5th

on the public leaderboard. Resolved lookalike business records across 1.7M entities with LoRA-tuned Qwen judges.

Papers

  1. 2027
    Seeing but Not Speaking: Most Token-Mass Interference in Budget-Bounded Video Question Answering Is a Decoding Artifact

    A. Verma, A. Rawat. WACV 2027, Applications Track.

    arXiv incomingOpenReview incoming

    Accuracy versus visual tokens per frame for short and long video benchmarks
  2. 2026
    COSMOS: Soft Mechanism Mixtures with Verifiable Routing for Long-Horizon PDE Forecasting

    A. Rawat, M. Padmanaban, J. Hazra. NeurIPS 2026 Workshop (STODY), poster.

    arXivOpenReview

    COSMOS architecture: encoder, four physics experts, soft router, fusion and decoder
  3. 2026
    GeoNLI: A Natural Language Interpreter for Satellite Imagery

    A. Gandhe, A. Rawat, G. Sethi, et al. Inter IIT Tech Meet 14.0, ISRO.

    arXiv

    GeoNLI pipeline: captioning, VQA and grounding models with majority voting

All papers on Google Scholar ↗

Work

  1. 2026–

    Conscious Engines

    Research Engineer Intern

    Draft-and-judge decoding for the SDAR-30B diffusion MoE: 1.35× end-to-end speedup.

  2. 2026

    IBM Research

    AI Research Intern

    Mixture-of-experts neural operator for mixed physics. Became COSMOS.

  3. 2026

    Whettle

    Technical Lead

    Sole engineer of a hotel operating system running at live properties.

  4. 2025

    Wyzr Fintech, London

    Data Science Intern

    Rule-first, LLM-fallback reconciliation; +33% match rate.

Hackathons

  1. 2026

    Hyde AI Builder Hackathon

    IIT Bombay, team catana

    A Catan agent on one CPU core. Won 67% of games against the reference bot.Event

    A Catan board mid-game: settlements, cities and roads on the hex tiles
    Game AIBeam searchHeuristic evaluationMLP ensemblesCPU-bound inference
  2. 2026

    Amazon ML Challenge

    5th, public leaderboard

    Matched 1.7M businesses across three noisy sources: retrieval, LoRA-tuned Qwen judges, LightGBM.Code

    Entity-resolution cascade: dense retrieval, cross-encoders and a LightGBM matcher for clear cases, eleven LoRA judges on Qwen3.5-4B and a LightGBM combiner for uncertain ones, and an unseen-country matcher for France
    Entity resolutionDense retrievalLoRA fine-tuningLLM judgesLightGBM
  3. 2026

    Gemini Hackathon

    Ghar, Live API track

    Talk to a live AI designer, get room concepts, narrated tours and an architect's build pack.Deployed website · Video · Code

    Ghar tour frame: an LED-lit living and dining room designed by the AI
    Agentic workflowsTool callingGemini Live APIReal-time voiceImage generationVideo generationSearch grounding
  4. 2026

    Amazon HackOn S6

    ReVival

    Seven agents that grade, price and resell returned products instead of sending them to landfill.Deployed website · Code

    ReVival storefront: Second Life, First Choice, certified refurbished products
    Agent orchestrationMulti-agent systemsVision modelsDynamic pricingAWS BedrockCircular commerce
  5. 2025

    Inter IIT Tech Meet 14.0

    ISRO problem statement

    GeoNLI, a natural-language interpreter for satellite imagery. Became a paper.Paper

    GeoNLI pipeline: captioning, VQA and grounding models with majority voting
    Vision-language modelsRemote sensingVQAVisual groundingDiffusion modelsSAR to optical
  6. 2025

    Kaggle super-resolution

    Rank 4

    4× upscaling for game frames that beat RCAN and SwinIR baselines.

    Super-resolutionComputer vision
  7. 2024

    Qualcomm VisionX

    Rank 4 of 150+ teams

    Reflection removal from a single photo, built to hold up under adversarial inputs.Code

    Reflection-removal architecture: multi-scale transformer blocks with cross-scale attention, multi-scale fusion and an adversarial image discriminator
    Reflection removalVision transformersAdversarial robustness

Projects

  1. 2026

    MacToolKit

    Native macOS app In development

    Window snapping, a window switcher with live previews, clipboard history and a launcher, in one menu-bar app.

    Diagram of MacToolKit snapping windows into a half-and-quarters layout
    macOSSwiftSwiftUIAppKit
  2. 2026

    PriorityKV

    With Arush Sharma

    KV-cache compression for agents that keeps what each message is for: tool schemas stay intact at 18× over budget.Code

    PriorityKV eviction and mixed-precision page paths
    LLM inferenceKV-cache compressionAgentsMixed precisionLong context
  3. 2026

    AlignmentLab

    With Arush Sharma

    Matched-compute GRPO on Qwen3-8B: rewards that ignore correctness drop GSM8K from 0.86 to about 0.5.Code

    Qwen3-8B accuracy on GSM8K and MATH-500 under random and format-only rewards
    Reinforcement learningGRPOReward designLLM post-training
  4. 2026

    VID-CURATOR

    WACV 2027, with Aman Verma

    Long-video QA on a frame budget: 8 curated frames beat 16 evenly spaced ones on NExT-GQA.

    Accuracy versus visual tokens per frame for short and long video benchmarks
    Video understandingVision-language modelsFrame selectionLong-video QA
  5. 2026

    Linguistically Guided MTP

    CS 769 In progress

    Is multi-token prediction harder in Hindi because of the tokenizer or the grammar? Romanised Hindi separates the two.

    Diagram: byte-level and subword sequences, predicting the next word group in English and Hindi
    Multi-token predictionTokenizationMultilingual NLP
  6. 2025

    Medusa decoding

    CS 726

    Multi-head decoding on Medusa Vicuna-7B: 2.3× faster than single-head decoding.Code · Medusa paper

    Medusa architecture from Cai et al., 2024: extra decoding heads on the last hidden state propose tokens for the next positions, verified in one step
    Speculative decodingLLM inferenceMulti-head decoding

Academics

Education

Scholastic achievements

  1. 2021

    KVPY Fellowship

    Government of India

    Top 1,000 in India.

  2. 2021

    Indian Astronomy Olympiad

    Stage 2

    Top 255 of 10,000+.

  3. 2019–20

    PRMO

    Qualified twice

    Top 300 in North-West India.

Teaching

  1. 2026

    Teaching Assistant

    Intro to AI/ML, IIT Bombay

    90+ students.

  2. 2024

    Mentor

    Seasons of Code, WnCC

    21 students, time-series forecasting.