Portfolio

Open-source autonomous computer agent with multi-agent orchestration, browser automation, persistent memory, sandboxed code execution, and workflow automation. Inspired by Perplexity Computer with context engineering from Superpowers and GSD.

End-to-end perception pipeline for autonomous driving — BEV fusion from multi-camera inputs, PointPillars 3D object detection, lane segmentation with polynomial fitting, Kalman-Hungarian multi-object tracking, and occupancy grid generation.

Neurosymbolic vision framework combining deep perception (CLIP, YOLO, SAM) with Z3-based symbolic reasoning. Features differentiable symbolic grounding, concept bottleneck models, visual QA, and a solar panel defect detection pipeline for drone imagery.

Production-grade MLOps platform with experiment tracking, model registry with versioning and stage promotion, data drift detection, FastAPI model serving with A/B testing (canary, shadow, blue/green), CI/CD pipeline templates, and a Streamlit monitoring dashboard.

Lightweight, pip-installable ML monitoring SDK. Statistical drift tests (PSI, KS, Jensen-Shannon, MMD), streaming concept drift detectors (ADWIN, Page-Hinkley, DDM), SHAP-based feature attribution tracking, alerting integrations, and HTML report generation.

Multi-scenario simulation benchmark for autonomous vehicle decision-making. Configurable scenarios (weather, traffic, edge cases), safety metrics (TTC, PET, jerk), four planner implementations (Lattice, RRT*, MPC, RL), behavior trees for NPC traffic, and a leaderboard system.