Research, projects,
and shipped systems.

av-policy-lab

2026 – present · MS research, Northeastern

Do diffusion planners beat a capacity-matched deterministic planner in nuPlan closed-loop, and does the gap grow where scenes are interaction-critical? A pre-registered 2×2 (head × goal type, 3 seeds, 800 scenarios, two traffic modes) says no, and the write-up explains why: under standard imitation the diffusion policy collapses to a near-deterministic point estimate, and the standard closed-loop metrics are heavily saturated. The follow-up locates the real bottleneck: a 6-mode policy already contains better maneuvers than the deterministic head, but its own mode choice is no better than random, and the gap widens under learned SMART reactive agents. Wayformer-style encoder and DDPM / DDIM heads built from scratch in PyTorch, trained on A100s via SLURM.

• best-of-modes beats deterministic: +0.144 ± 0.025 CLS, 3 seeds
• the policy's own mode pick: +0.007, no better than random
• diffusion collapse: median 0.47 m from deterministic

view on github → read the write-up →

Laksh.ai

2025 – present

Upload a jump-shot video and get a scout report. MediaPipe Pose tracks the shooter in 3D, a physics pipeline turns the motion into 8 biomechanical metrics (release velocity, arc, knee and elbow angle, kinetic sync, fluidity, hip rotation, balance), ChromaDB finds the nearest of about 500 active NBA player profiles (8D vectors derived from box-score stats) by cosine similarity, and Gemini 2.5 Flash writes the report and three drills. Presented at the YC × Google DeepMind hackathon; won a prize at the AI Prototype to Product hackathon.

• 8D kinematic fingerprint from 3D pose
• nearest match among ~500 NBA profiles
• FastAPI + React app, Dockerized

view on github →

Vision Transformer from scratch

2025 – 2026

Patch embedding, multi-head self-attention, and transformer blocks written by hand, no pretrained weights. On CIFAR-10, 12 ablations at a matched ~1.21M parameters show which parts carry the model: removing residual connections costs 30.6 points, Post-Norm adds 2.4. On LEVIR-CD change detection, a Siamese ViT, U-Net, and Swin share one training pipeline. The CNN wins, which fits the idea that local inductive bias still matters on high-resolution, imbalanced masks.

• no residuals: 78.3% → 47.6% on CIFAR-10
• LEVIR-CD F1: U-Net 0.884 · Swin 0.861 · ViT 0.824
• best epoch by validation F1, from the logs

view on github →

Achilles tendon load

2026

A feasibility study: can a cheap insole-style signal recover the load inside the Achilles tendon, the quantity that drives injury? Tendon force (ankle moment over moment arm), stress, and strain are computed on two open BMClab datasets and cross-checked against OpenSim, which agrees within about 16%; a full OpenSim pipeline (scale, IK, inverse dynamics) runs on a second walking dataset. Models then predict the whole force curve from six wearable-style inputs, scored with 5-fold cross-validation that holds out whole people.

• 73 people, 839 walking + running step-cycles
• R² 0.98 vs a 0.91 no-skill floor (0.94 loaded phase)
• linear model recommended: 7.5% peak-force error

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

An open-source Model Context Protocol connector, wired into Claude, Cursor, and ChatGPT, that asks you to explain what you just learned. It grounds the check in your own sources, scores the explanation against verified evidence instead of letting you grade yourself, and resurfaces what you do not know yet. Local-first, no telemetry.

github →

Trace

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

github →

TaskRoute

A FastAPI routing service that scores provider offerings (LLM APIs and compute services) on cost, latency, quality, reliability, and switching cost, then ranks them against a task's objective and budget.

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 →