Projects

Selected projects I’ve built, each with the stack and the hard parts. Write-ups live on the blog; code lives on GitHub.

inference-bench

Python · vLLM · SGLang · TGI · asyncio · benchmarking

Benchmark and compare LLM serving engines on a single GPU: vLLM vs SGLang (Llama-3-8B, RTX 4090, fp16). Full write-up on the blog.
Reproducible harness: per-engine virtualenvs (no dependency collisions), semaphore-gated concurrency, per-request JSONL + aggregated summaries. 960 requests/engine, 0% errors.
Key result: both engines are effectively tied on throughput (~56-60 tok/s at c=1, ~1000-1080 tok/s at c=32 via continuous batching). Throughput scales ~18x while latency stays flat.
Methodology rigor: fixed 128-token outputs (prompt engineered to avoid early EOS) and throughput measured as tokens/cell-wall-time, not tokens/slowest-request.

Python · PyTorch · Grounding DINO · SAM 2 · Depth Anything V2 · ONNX Runtime

Open-vocabulary 3D auto-labeling pipeline for autonomous-vehicle perception: generates KITTI-format 3D bounding boxes from raw dashcam video or nuScenes scenes, with no manual annotation.
Four-stage architecture: Grounding DINO for open-vocabulary 2D detection, SAM 2 for temporal mask propagation, Depth Anything V2 for monocular depth, and PCA-based 3D box fitting with metric-scale recovery.
Exported the depth model to ONNX with full graph optimization (constant folding, operator fusion, shape inference), achieving a 1.5-3x inference speedup on CoreML and CUDA.

Python · FastAPI · React.js · Mistral AI · ElevenLabs

AI-powered Cold War spy-agency simulator, built at the Mistral AI Worldwide Hackathon 2026 (Tokyo). The player directs a covert intelligence network across Tehran, Kyiv, Hong Kong.
Multi-agent architecture where each operative runs private autonomous reasoning (hidden from the player and the orchestrator), simulating real intelligence fog-of-war: secret compliance, deception, defection.
A rogue engine autonomously triggers defections from dynamic loyalty scores; ElevenLabs delivers real-time voice transmissions per operative.

Python · YOLOv8 · BoT-SORT · OpenCV · scikit-learn

End-to-end football vision analytics pipeline: broadcast footage in, track-level player intelligence, team-aware tactical views, and possession statistics out.
Fine-tuned YOLOv8s (4 classes) to mAP50 = 0.807 at ~53 FPS; BoT-SORT with global motion compensation cut player ID fragmentation from ~630 to 85 unique IDs.
Automatic team classification via HSV histograms + KMeans; pitch-keypoints pose model + RANSAC homography maps players into pitch coordinates for a synchronized top-down radar.

Python · BoT-SORT + ReID · homography · jersey OCR · Three.js / WebGL

Real-time 3D digital twin of a football match, reconstructed from plain 2D broadcast video: tactical top-down view, free-orbit camera, and a first-person view from any player.
Own CV pipeline for tracking and calibration; no proprietary league data required.

More on the way: fine-tuning experiments, serving-infrastructure notes, and more inference benchmarks.