A working archive of projects in computer vision, 3D perception, and the infrastructure that keeps them honest.
12 projects · 12 total
Real-time multi-camera person tracking for customer behavior analysis and site security, with re-identification across non-overlapping views. Pose triangulation and appearance embeddings raised inter-camera mAP from 40% to 74% on an open-set benchmark.
Low-latency face and license-plate anonymization for RTSP streams, integrated directly into Milestone VMS through the AI Bridge. Shared decoders and batching improved load balancing and kept GPU utilization high under concurrent camera load.
A real-time 3D reconstruction and contour-fitting pipeline using depth cameras and Open3D. Built as a multi-process pipe-and-filter system for airline cargo measurement at operational speeds.
Took over a 30,000-line monolith already running across 1,000+ stores in three EU countries, with undocumented workflows and mounting client complaints. Directed a seven-person team through a full audit, closing every open client complaint and securing a contract renewal. Improved the perception pipeline with SOTA models, migrated from centralized polling to an event-driven architecture that cut read traffic by 94%, and built platform tooling (automated deployments, rollback, and site management) that made scaling across 1,000+ stores operationally tractable.
Kubernetes-based model serving, benchmarking, and safety evaluation for autonomous vehicle ML products. Added new safety performance indicators and benchmark workflows for real and synthetically perturbed data.
A single training framework for classification, detection, and segmentation. Reworked CPU-heavy metrics with multiprocessing and Cython, pushing overall GPU utilization above 90% in training and inference.
A PostgreSQL and PostGIS platform for curated dataset storage, geographic querying, and sampling. Included a Django backend, Celery workers for heavy compute jobs, and a client library for dataset access.
A media-processing service for compression, metadata analysis, and beautification. Optimizations cut infrastructure cost by 70 percent while enabling plagiarism analysis and adaptive SSIM-based compression.
A high-performance DynamoDB-backed product form workflow that reduced payload size by 99.5 percent and improved form completion rate by 5x on mobile-heavy usage.
A clustering-based quality platform for catalog labeling review with human-in-the-loop analysis. Reduced catalog quality checks from weeks to under an hour.
A non-rigid registration-based atmospheric turbulence removal module for long-range imagery. CUDA and pipe-and-filter optimizations delivered a 240x speedup and made the system usable in real time.
A wide-area tracking system with PTZ control, multi-camera stitching, and change detection for CCD and infrared camera setups in real-time operational environments.
