Field-tested architectures for real-time edge computer vision: sub-15ms YOLO detection, multi-camera vehicle re-identification, Kalman filtering, and TensorRT acceleration processing 50,000+ daily events.
Building computer vision for production is entirely different from running Kaggle notebooks. Real-world edge systems must survive illumination shifts, camera vibrations, rain occlusion, bandwidth constraints, and thermal throttling on embedded hardware like the NVIDIA Jetson Orin. The case studies below document our production solutions.
18 min
Production detection layer under NHAI's 2023 ATMS Standards: multi-vendor ONVIF cameras, dual-model TensorRT inference, occlusion-resilient ByteTrack, 7 incident rules, and SHA-256 evidence integrity.
Shubham Kulkarni
14 min
Architecting an event-driven edge-to-cloud incident intelligence engine: sub-200ms anomaly validation, 6-agent LangGraph orchestrator, and real-time industrial SOP escalation.
Shubham Kulkarni
11 min
Stateless in-memory vehicle forensic matching using YOLO, OSNet-AIN metric learning embeddings, and FAISS IVF-PQ indexing across 150,000+ candidates in under 12ms.
Shubham Kulkarni
10 min
Architecting the VIDES computer vision system at Arya Omnitalk: live highway traffic management system processing 50,000+ daily events with YOLOv8 and DeepSORT.
Shubham Kulkarni
8 min
A comprehensive engineering comparison of anchor-based vs anchor-free architectures, NMS vs Hungarian matching, and TensorRT INT8 optimization tradeoffs.
Shubham Kulkarni
7 min
Deploying multispectral vegetation index analysis and soil telemetry on solar-powered edge nodes with automated LoRaWAN actuation.
Shubham Kulkarni