Deep Learning and Data Enthusiast
AI/ML engineer with strong focus on computer vision and video understanding, experienced in building end-to-end, production-oriented ML systems. Hands-on experience in action recognition, spatio-temporal modeling, dataset engineering (AVA-style formats), and deep learning using PyTorch and TensorFlow. Actively working on open-source and research-oriented projects involving video analytics, surveillance intelligence, and scalable ML pipelines. Comfortable owning the full lifecycle—from data curation and model training to deployment-ready tooling—and continuously strengthening foundations in DSA and system-level thinking to operate effectively at scale.
Deepnet Analytics
8 mos · New Delhi, Delhi, India
DeepNet Analytics
June 2025 - Jan 2026
Co-developed OpenAVA-Lab, an open-source AVA-style spatio-temporal dataset creation platform reviewed at AAAI 2025. Built an RF-DETR and KalmanSORT automated proposal pipeline at 25+ FPS, with 7.2x throughput gain and 87% reduction in active annotation time. Exposed orchestration through FastAPI REST APIs and deployed on AWS EC2, Lambda, and RDS using Docker and PostgreSQL. Implemented YOLOv9 PPE detection achieving 90% mAP on a custom 8-class industrial dataset.
Bachelor of Technology - BTech,Electronics and Communications Engineering
Jul 2023 - Jul 2027
High school
[object Object]
plus 2
[object Object]
B. Tech, Electronics & Communication Engineering
2023-2027
CGPA: 8.21/10. Co-author of OpenAVA-Lab, AVA-style Spatio-Temporal Dataset Platform (reviewed at AAAI 2025; positive reviewer feedback).
Startup MVP. Built an AI-powered PPE compliance system achieving 95% mAP across 8 safety-equipment classes with end-to-end inference latency under 60 ms. Processed 10,000+ annotated frames and architected an event-driven Kafka alert system for concurrent camera streams. Deployed Dockerized microservices on AWS GPU instances.
Built a LangGraph multi-agent analytics platform with a RAG pipeline achieving 80% query accuracy and an asynchronous FastAPI REST layer for conversational business-metrics queries.
Built an end-to-end 4-stage cascading presentation-attack-detection security waterfall achieving 95.2% accuracy and 0% false-positive rate on a mixed-domain evaluation set. Designed a Frequency-Temporal Cross-Attention architecture and implemented CHROM rPPG liveness, PRNU sensor-noise forensics, and Moiré FFT replay detection.