Shrikar Madhu
Embodied AI · Computer Vision · World Models
I'm a Graduate Researcher at NYU Courant, working with Prof. Saining Xie on long-context video understanding, efficient sequence modeling, and spatial reasoning in vision systems.
My broader interests sit at the intersection of computer vision, continual learning, and spatial supersensing for embodied AI — building visual systems that perceive, reason about, and act in physical environments.
Before NYU, I was a Pre-Doctoral Fellow at the Kotak IISc AI-ML Research Centre, where I was advised by Prof. Suresh Sundaram and published at CVPR 2025. I also co-founded Physi.Fit — an idea that started with my own ACL injury and became a funded AI healthcare platform.
Affiliations
Publications & Patents
VISTA-CLIP: Visual Incremental Self-Tuned Adaptation for Efficient Continual Panoptic Segmentation
Continual learning framework using visual prompt tuning (<1M parameters) for panoptic segmentation in autonomous driving. Achieves state-of-the-art 62.3 mIoU on Cityscapes while preventing 53% catastrophic forgetting.
IndraEye: Infrared Electro-Optical UAV-based Perception Dataset
First multi-sensor, multi-domain slant-angle EO-IR dataset (50K+ aligned pairs) for UAV perception. Addresses occlusion and scale challenges in aerial object detection and semantic segmentation. Pix2Pix GAN cross-modal synthesis achieves 0.85 SSIM, improving all-weather perception by 28%.
Detection of Conversational Health in a Multimodal Conversation Graph
Graph labeling framework for ranking multimodal conversations via emotional concordance scoring. Uses late fusion in graph attention networks to learn node importance through representation learning — outperforming prior state-of-the-art.
Method and System for Conducting Interactive Rehabilitation Sessions with Continuous Monitoring
Patented deep learning framework fusing video-based pose estimation with multi-modal biomarkers via attention-based GNNs, for real-time orthopedic rehabilitation monitoring. Deployed across 17 healthcare facilities.
Experience
Graduate Researcher NYU Computer Vision Lab · Prof. Saining Xie
- Built Pico-LLM (Gated Linear Attention) achieving O(N) complexity, outperforming Transformers 8–16× on 5k-token sequences with 94% reasoning accuracy via GRPO-based RL post-training.
- Long-context video understanding system on 8× A100 GPUs with custom CUDA kernels: 7% mAP improvement, 40% memory reduction over baseline.
- Sparse transformer spatial reasoning: 50+ ablation experiments across 3 benchmarks, 2.1× inference speedup.
Pre-Doctoral Scholar Kotak IISc AI-ML Research Centre
- Developed VISTA-CLIP for continual panoptic segmentation — published CVPR 2025. SOTA 62.3 mIoU with visual prompt tuning, 53% less catastrophic forgetting.
- Edge deployment with selective parameter adaptation and knowledge distillation: 3.2× lower latency at 98% accuracy on embedded devices.
- Curated first multi-sensor EO-IR slant-angle aerial perception dataset (50K+ pairs) — submitted ICRA 2025.
Co-Founder & AI Research Engineer Physi.Fit, Inc.
- Built from a personal ACL injury into a $250K seed-funded (100X.VC) platform deployed across 17 facilities serving 100+ patients. Acquired by a Khosla Ventures-backed healthcare company.
- 30 FPS real-time pose estimation pipeline (Mobile MediaPipe, CNNs, Video-LLaVA) across 10K+ sessions — 92% motion accuracy, 25% reduction in therapy time.
- Granted U.S. patent for deep learning framework for continuous monitoring in rehabilitation.
Research Assistant IISc AI & Robotics Lab
- Hybrid spiking neural network (SEFRON + Temporal-SNN) with STDP learning: +4.35% over CNN on CIFAR-10 at 60% lower energy — targeting neuromorphic edge AI applications.