AV-Policy-Lab
Closed-loop reinforcement and imitation learning for autonomous driving in nuPlan simulation. Developing behavior cloning, a MILE-style world model, and VLA-conditioned policy baselines evaluated on PDM-Score metrics. DDPM and DDIM built from scratch in PyTorch. Training loss from 0.75 to 0.032, 100x inference speedup with deterministic sampling. Running on HPC A100 via SLURM.
nuPlan SLURM A100 DDPM DDIM World Model VLA
Laksh.ai
Won the AI Prototype to Product Hackathon (March 2026). Production biomechanical pose estimation pipeline for real-time sports analysis. Benchmarked MediaPipe, RTMPose, and YOLOv10 on a custom motion dataset for 8D kinematic fingerprint extraction. Deployed via FastAPI with ChromaDB vector search for athlete similarity matching.
RTMPose YOLOv10 MediaPipe ChromaDB Vector Search
MediTrack
Healthcare equipment tracking system built at the Memories.ai Hackathon (November 2025). Natural language queries over hospital surveillance footage using Memories.ai's Large Visual Memory Model, with real-time utilization analytics and configurable alerting across equipment types.
Supabase LVMM NL Search Real-time Analytics
TaskRoute
Multi-agent LLM routing system that dynamically selects providers based on task type, cost constraints, and reliability signals. Routing decisions served through a FastAPI decision API.
Multi-Agent LLM Orchestration Provider Routing
ViT from Scratch
Vision Transformer built from scratch in PyTorch for bi-temporal change detection on Sentinel-2 satellite imagery. Patch embedding, multi-head self-attention, and transformer encoder blocks with no pretrained weights. Benchmarked across IoU, F1, Dice, precision, and recall to map architecture tradeoffs.
Transformers Sentinel-2 Remote Sensing Change Detection
Background
Mechanical engineering at NTU Singapore. AV test engineer at Venti Technologies, testing perception and navigation systems in live port environments using RViz and LiDAR, cutting testing cycle time by 15%. Site reliability engineer at United Overseas Bank on distributed banking infrastructure on AWS. Graduate TA for Foundations of AI at Northeastern, supporting 30+ MS students on deep learning and NLP projects.
Now year one of my MS. Focused on generative models and physical AI. Next step is deploying Diffusion Policy on an SO-101 robot.


