I love ML and doing hard stuff.
Fond of burning GPUs.
Senior at IIT Roorkee.
Machine Learning Intern | Media & Data Science Research (MDSR) Lab
Distributed training pipelines for RL fine-tuning multimodal LLMs (InternVL3 8B, Qwen2.5VL) using GRPO. Beat openai-o3 on set of high-resolution image tasks requiring fine-grained visual analysis.
Also trained AutoShot derivative models on ad campaign videos for scene segmentation and some more stuff with ad videos.
Research Collaborator | Digital Twin Industries
Trained a conditional 1D diffusion model and DDIM-inversion editing pipeline to generate/modify normal and faulty vibrations. Achieved SOTA results with limited data. Also trained a 1D ViT classifier that improved accuracy by 30% over previous methods on held-out sets.
Ved Umrajkar*, Aakash Kumar Singh*
InterIIT Tech 13 — 4th Place
Built a multimodal RAG system to query real-time world state, allocate tasks, and perform path planning. Developed ROS2 simulation in Gazebo integrating continuous mapping, exploration, and navigation using only visual & depth perception.
Designed supervised contrastive learning for fundus image severity classification and case-based retrieval. Significantly improved mAP and AUC over transfer learning baselines; integrated HNSW for efficient retrieval.
Implemented quantized 8-bit CNN inference on Basys3 FPGA for hardware-accelerated real-time image classification.
Secured 16th rank nationally out of 10,000+ teams.
Silver Medal (Rank 55/1161).