ML Researcher · NIT Kurukshetra · Class of 2027

Efficient models
for real data.

I work on state-space models, hyperspectral vision, and efficient sequence modeling — building systems that are both mathematically principled and practically deployable on real hardware.

Currently collaborating with two universities across continents, interning at a startup, and building my own company.

About me

I'm a second-year undergraduate at NIT Kurukshetra working on applied machine learning research. My interests are efficiency and robustness — I believe in models that are surgically precise, ones that understand underlying structure rather than memorize surface statistics.

This philosophy drives everything from LightMedSeg (3D medical segmentation in 0.48M params) to SCAF (hyperspectral classification via optimal transport). I'm interested in work where mathematical elegance translates to practical performance.

Beyond research I'm building Reflara, an execution-intelligence platform for autonomous AI agents, and shipping agentic pipelines at Qten.ai. I've contributed 25+ merged pull requests to Lightning AI and OpenCV. I'm also an Operations Team Member at SIGPLAN-M, organizing international mentoring for programming languages researchers.

Currently B.Tech ECE, NIT Kurukshetra — CGPA 8.18
JEE Mains 2023 98.54 percentile
Class XII · X 92% · 96%
Collaborations Sapienza University of Rome · Univ. of Tübingen
Industry Qten.ai · Founder @ Reflara
Open Source 25+ merged PRs (Lightning AI, OpenCV)
Interests SSMs · Optimal transport · Hyperspectral vision · Formal methods

Research experience

Four labs.
One mission.

Efficient, robust models that work on real hardware under real constraints. Each project approaches the same core challenge from a different angle.

01

Sapienza University of Rome · May 2026 – Present · with Paulo Russo

Physically Grounded SSMs for Hyperspectral Super-Resolution

Proposed AbSpec-Mamba and VolMamba v2 — models that condition selective state transitions on per-pixel material abundance, using prototype-guided reconstruction to recover fine spectral detail without hallucinating physically implausible spectral signatures.

Building content-adaptive multidimensional scanning and progressive coarse-to-fine reconstruction so the model restores high-resolution hyperspectral images under severe sensor degradation and spectral corruption.

Active
02

University of Tübingen · Sep 2025 – Present · with Andreas Zeigler

CE-GNN: Causally Constrained Graph Networks for Event-Based Vision

Formally characterized how leading event-vision GNNs (AEGNN, SlideGCN, Voxel-GNN, HUGNet2+PA) leak future information backward in time, violating the causality required for true real-time inference on event-camera streams.

Designing CE-GNN to enforce strict causal constraints via O(k log N) incremental graph construction, directed temporal attention, and predictive edge insertion. Manuscript in preparation.

Manuscript prep
03

IIT Ropar · May – Aug 2025 · with Dr. Puneet Goyal

SCAF: Robust Hyperspectral Classification via Optimal Transport

Identified that standard HSI pipelines discard recoverable inter-band correlation by treating bands independently. Recovered this signal using differentiable optimal-transport grouping with balanced Sinkhorn assignments.

Per-group Mamba processors isolate sensor noise to individual groups, converting spectral redundancy into a graceful-degradation mechanism — achieving +11.6% accuracy under 20–30% band dropout against Transformer baselines.

Under review
04

NIT Kurukshetra · Sep 2024 – May 2025 · with Dr. Vishwas Rathi & Dr. Shweta Sharma

3D Medical Segmentation & Malware Analysis

Built parameter-efficient 3D segmentation models targeting deployability on memory-constrained hardware, work that became LightMedSeg and RefineFormer3D. Concurrently developed hybrid frameworks for ransomware detection combining static binary features with NLP-based code representations, leading to the ICDAM 2025 paper.

2 papers

What I'm building

Startups and
systems that ship.

Research is one side. The other is building things people actually use.

AI Engineer Intern · May 2026 – Present

Qten.ai

Building agentic orchestration for grounded conversational content generation at scale. Multi-layer pipelines that verify every claim against retrieved evidence and self-repair low-quality output.

  • Multi-layer generation pipelines with narrative constraints and deterministic evaluators
  • Modular LLM routing dispatching to reasoning agents over structured latent planning
  • Cuts cost and latency while sustaining high-engagement content generation at scale

qten.ai →

Selected project · Jan 2026 – Present

S6MOD: Adaptive Mixture-of-Experts

Replaced fixed routing temperature with a regret-driven adaptive controller, letting the model tune its own exploration vs. exploitation as conditions shift.

  • Reformulated routing as closed-loop feedback driven by loss trends and expert-usage entropy
  • Improved stability in online continual learning under distribution shift
  • Minimal architectural changes — drops straight into S6MOD

Publications & preprints

Four papers.
All first-author.

LightMedSeg: Lightweight 3D Medical Image Segmentation with Learned Spatial Anchors

Tyagi et al. · CVPR Workshops, Efficient Computer Vision (ECV 2026)

84.8% Dice on BraTS · 93.5% on ACDC · 0.48M parameters · 14 GFLOPs. Parameter-efficient 3D segmentation for memory-constrained clinical hardware with strong label efficiency.

CVPR Workshop 2026

SCAF: Robust Hyperspectral Classification via Optimal Transport-Based Spectral Grouping

Tyagi et al. · Preprint, under review

Differentiable optimal-transport grouping with balanced Sinkhorn assignments. Mamba backbone cuts complexity from quadratic to linear. +11.6% accuracy under band dropout.

Under review

RefineFormer3D: Hierarchical Multi-Scale Transformer for Parameter-Efficient 3D Medical Segmentation

Tyagi et al. · Preprint, under review

86% Dice on BraTS 2017 · Outperforms SegFormer3D in efficiency and label efficiency for low-annotation settings.

Under review

Ghidra-Assisted Static Analysis and Ensemble Learning with Differential-Privacy GANs for Ransomware Detection

Chaturvedi et al. · ICDAM 2025 (Accepted)

Hybrid framework combining static binary features with NLP-based code representations and differential-privacy GANs.

ICDAM 2025

Open source

Code shipped to
real codebases.

Not just forks. Actual merged work reviewed and adopted by core maintainers.

16 PRs merged

Lightning AI

Improved model compilation pipelines, fixed argument parsing edge-cases across training configurations, and hardened training stability under distributed settings. Changes reviewed and adopted by core maintainers.


github.com/Lightning-AI/lightning
9 PRs merged

OpenCV

Targeted bug fixes and API robustness improvements across core computer-vision pipelines. Hardened edge-case handling in critical image processing routines used by thousands of applications.


github.com/opencv/opencv

Technical skills

Research areas

State Space Models (Mamba, S6) · Causal & Event-Based GNNs · Hyperspectral Classification · 3D Medical Segmentation · Parameter-Efficient Vision Transformers · Optimal Transport

Frameworks

PyTorch (AMP, custom modules) · PyTorch Lightning · TensorFlow · Keras · Scikit-learn

Libraries

OpenCV · NiBabel · TorchIO · Albumentations · NumPy · Pandas · Matplotlib

Tooling

Git · Docker · CUDA · Google Cloud Platform · Linux · Ghidra (Jython) · Jupyter

Languages

Python · C++ · SQL · LaTeX · Jython

Get in touch

Let's collaborate.
Research or startup.

I'm always interested in conversations about research, collaborations, or ideas worth pursuing. Drop me a message — I read everything.