Available for freelance & full-time opportunities
I build production-grade computer vision systems — from real-time face recognition at 70 lakh+ identity scale to person re-identification models running on Raspberry Pi at the edge. 4+ years turning research into deployed, reliable AI.
About Me
I'm a Computer Vision & AI Engineer at Magellanic Cloud Ltd (previously IVIS International), building real-time video analytics systems used across banking, retail and surveillance deployments in production.
My work spans the full pipeline: training deep learning models, optimizing them with quantization (PTQ/QAT) and hardware acceleration (TensorRT, OpenVINO), and deploying them on everything from NVIDIA GPUs to ARM edge devices like the Raspberry Pi 5 — without sacrificing accuracy or latency.
I care about shipping AI that actually works in the field: fewer false alerts, faster inference, and systems that hold up across 100+ live site cameras.
What I Work With
Career Path
Magellanic Cloud Ltd, Hyderabad
IVIS International Private Limited, Hyderabad
Selected Work
Real-time customer identification from live CCTV across 70 lakh+ enrolled identities using ArcFace + RetinaFace, with HNSW approximate nearest-neighbor search and TensorRT inference. Delivered as a C++/JNI library for Java banking applications.
Trained a ResNet-50 + ArcFace head with Batch-Hard Triplet Loss from scratch, quantized to INT8, and deployed live at 4 FPS on Raspberry Pi across bank and retail sites via OpenVINO.
A novel, from-scratch keypoint-based system that measures coverage between guard hands and employee body keypoints to flag incomplete security frisking (<90%) in real time, replacing manual CCTV review.
Productizing surveillance AI as deployable microservices: RTSP ingestion → YOLOv8 detection → Redis state → FastAPI → WebSocket dashboard, deployed on AWS.
View RepoObject tracking and re-identification pipeline for accurate gate-level people counting, with real-time event visualization through a lightweight web client.
A web tool to run and monitor AI inference on server-side video data, with interactive selection of models, classes, sites, and date ranges.
A dedicated model to filter good vs. bad quality face captures at registration time, boosting downstream face-recognition accuracy.
Custom motion-triggered object detection built on a quantized Int8 IR person-vehicle model for efficient, always-on site monitoring.
Camera tampering detection, gender detection, vehicle detection, weapon detection, object color prediction, no-video detection, image quality enhancement, and staff detection — shipped across live customer sites.
Background
Master of Computer Application · July 2018 – June 2021
Operating Systems, Data Structures, Analysis of Algorithms, Artificial Intelligence, Machine Learning, Networking, Databases
Get In Touch
Have a project, role, or idea in mind? My inbox is open — I usually reply within a day.