I build agents that plan ahead.

A fully vectorized PyTorch planning agent — ~5,500 LOC across 15 modules — that played thousands of live 500-turn adversarial ladder games inside a hard per-turn compute budget.

• Built an orbital trajectory-prediction engine that forward-simulates planet motion over an 18-turn horizon and solves intercept angles against planets rotating around a central sun.
• Paired it with an ETA-aware reinforcement-risk model that declines captures the opponent could reinforce mid-flight.
• Depth-2 expectimax1 with softmax opponent modeling for 1v1; a separately tuned game-theoretic policy for 4-player free-for-all.
• Prototyped behavioural-cloning and reinforcement-learning agents, benchmarked all three head-to-head on the ladder, and shipped the strongest.

An end-to-end 3D scene-reconstruction pipeline from a single 360° panorama — monocular depth estimation, object detection, and segmentation, orchestrated into one pass.

• Reached 91% segmentation accuracy across indoor and outdoor scenes.
• Converted equirectangular panoramas to cube-map projections and lifted 2D detections into 3D with depth models — 23% lower depth-estimation error than off-the-shelf baselines.
• Fully terminal-driven and reproducible: arbitrary input panoramas, no manual intervention between stages, under 90 seconds per scene on standard GPU hardware.

A resume-parsing and job-matching system built on BERT fine-tuning — 94% skill-extraction accuracy, benchmarked across 5+ training configurations.

• Improved recommendation precision by 18% over a TF-IDF baseline.
• Containerized a dual-interface web app with Docker and exposed inference via REST APIs; cut median latency 40% with batched scoring and caching.
• Automated bulk ingestion and candidate ranking — 500+ resumes processed end-to-end, manual screening time down 70%.