Research, projects,
and shipped systems.

av-policy-lab

2026 – present · MS research, Northeastern

A controlled study of when generative (diffusion) planners actually beat deterministic ones on nuPlan closed-loop, and whether that advantage survives once logged playback is replaced with realistic reactive agents. Baselines span behavior cloning, a MILE-style world model, and a VLA-conditioned policy, scored on PDM-Score; DDPM and DDIM are implemented from scratch in PyTorch and trained on A100s via SLURM.

• training loss 0.75 → 0.032
• ~100× faster sampling with deterministic DDIM
• honest failure-mode analysis + reusable harness

view on github →

Laksh.ai

2025 – present

A computer-vision pipeline for basketball jump-shot analysis: a pluggable pose stack (MediaPipe + RTMPose) feeds a canonical 12-joint schema, an 8D kinematic fingerprint is matched against active NBA players via ChromaDB vector search, and an LLM returns coaching feedback. Presented at the YC × Google DeepMind hackathon; won a prize at the AI Prototype to Product hackathon.

• canonical 12-joint schema → 8D fingerprint
• vector match vs active NBA players
• injury-risk: knee valgus, landing, upper-extremity load

view on github →

Vision Transformer from scratch

2025 – present

A ViT built from scratch in PyTorch for bi-temporal urban change detection on Sentinel-2 imagery: patch embedding, positional encoding, multi-head self-attention, and a Siamese encoder with a segmentation decoder, no pretrained modules. Evaluated honestly against tuned CNN baselines on IoU, F1, Dice, precision, and recall.

• ViT: F1 0.82, IoU 0.70
• tuned baselines: U-Net 0.88, Swin 0.86 (F1)
• a clear, reported gap, not a cherry-picked win

view on github →

career-ops

An AI-powered job-search and application system on Claude Code: 14 skill modes, a Go dashboard, automated PDF generation, batch processing.

github →

Feynman-Loop

A tool for thought that makes you understand concepts deeply: you explain a concept, it finds the gap against a trusted source and resurfaces what you don't yet truly know. Built as an MCP server.

github →

Trace

A multi-LLM newsletter agent that infers what you are curious about and assembles a personalized newsletter around those interests.

github →

TaskRoute

An LLM-provider optimization engine ranking providers by task type, budget, switching cost, and reliability risk in a simulated agent marketplace.

github →

MediTrack

Healthcare-equipment tracking using visual-memory AI for semantic search over surveillance footage. Memories.ai Hackathon.

github →

CRPO

Contrastive Reasoning Prompt Optimization: prompts robust across math, reasoning, fact verification, and code generation.

github →