Applied Scientist & AI Researcher specializing in
Agentic AI, Multimodal LLMs, and Computer Vision.
M.Tech, AI @ IIT Delhi · ICLR 2027 submission · Top 10 in Amazon ML Challenge 2026 · Publications at CVPR, MICCAI & ICASSP.
Building AI systems from research prototypes to production-scale deployments.
From faster multi-agent LLM systems and entity resolution at scale to medical imaging, depth estimation and robotics.
IIT Delhi · Prof. Sayan Ranu
Lossless speculative execution for multi-agent LLM systems. Predicts which agent runs next (link prediction over an agent graph mined from past traces, with a lightweight GRU) and what it receives (a speculator LLM with retrieval-augmented in-context learning), then runs it early and keeps only correct results.
Amazon ML Challenge 2026
Built in 3 days a pipeline linking 9.97M noisy multilingual records to 1.73M entities across 3 countries: a fine-tuned e5 bi-encoder with FAISS HNSW for blocking, cross-encoders and two-stage LightGBM for matching, an LLM judge for uncertain pairs, and self-training for an unlabeled country.
University of Pennsylvania
Novel segmentation algorithm using modified nnU-Net with dual-encoder design, cross-attention, ODConv3D layers, and multi-scale attention strategy.
KAUST
Adapted ZoeDepth for panorama 360° images with spline optimization, pixel-wise adjustment networks, LoRA adapters, and perceptual loss functions.
IIT Jammu · Prof. Vinit Jakhetiya
Fine-tuned DepthAnything with Mask Loss, Edge Loss, L1 & SSIM Loss for transparent/mirror surface depth estimation.
IIT Jammu · Prof. Vinit Jakhetiya
Retrained NAFSSR on DIV2K with perceptual/texture loss. Achieved SSIM 0.7268. Fine-tuned NAFSSR-T_x4 with FP16 to sub-1 MB model size with ODConv2D layers.
IIT Jammu · Prof. Vinit Jakhetiya
YOLOv7 + CSRT tracker pipeline with post-processing to minimize false positives for drone detection and tracking.
Flipkart GRiD 3.0
Centralized parcel sortation system managing 4 sensorless bots via single camera feed with Aruco marker localization and collision avoidance.
IIT Jammu
Interfaced six stepper motors with machined parts via Teensy micro-controller. Applied Python-based inverse kinematics for precise grasping and depositing.
IIT Jammu
3D-printed human-like arm using servos, strings, Raspberry Pi, and camera to replicate real-time human finger positions via ResNet18.
A paper under review at ICLR 2027 and 4 publications at top-tier venues in computer vision, medical imaging and signal processing.
Ranked in the top 10 of 32,000+ teams in business entity resolution over ~10M multilingual records.
Reviewed scientific manuscripts for the MICCAI BrainLesion Workshop 2023.
Presented a paper at the IEEE ICASSP 2023 Conference in Rhodes, Greece. Hosted a "Demystifying Neural Networks" session for deeplearning.ai.
Secured 1st rank at the MICCAI 2023 BraTS Africa Challenge (brain tumor segmentation).
2nd place in the National Finale of the Flipkart GRiD 3.0 Robotics Challenge.
2nd rank in Drone-vs-Bird (ICASSP 2023), Mind Maze & Auto-BOT at TRYST (IIT Delhi). 3rd in Pitch for Better India.
Volunteered as Residence Management Head at CVIP Conference 2023.
Looking for 2027 research and applied scientist internships: available for 4–6 weeks in December 2026, or 12–14 weeks from mid-May 2027. Also open to research collaborations. Feel free to reach out!