About

PyTorch · YOLO · PARSeq · DeepStream · Triton · TensorRT · FastAPI · Kubernetes · Docker · Bangalore

• ·Shipped a 10+ module CV system (LPR, PUCC, helmet, phone, wrong-way) over live RTSP at 0.9+ F1 on Indian CCTV.
• ·Wrote a custom-C DeepStream pipeline that keeps decode → detect → track on-GPU, removing CPU↔GPU copies. Scaled per-server capacity to 100+ cameras on multi-GPU.
• ·Cut detection latency 54% to ≤10 ms by replacing anchor-based heads with NMS-free YOLOv10 on Triton + TensorRT. 0.90+ mAP across 8 classes.
• ·PUCC pollution-certificate FastAPI service: YOLO verifies probe insertion, LPR reads the plate, OCR'd testing-center name validated against the government registry.
• ·Rebuilt LPR end-to-end: YOLOv11s for text detection, PARSeq ViT for recognition, TensorRT FP16 on edge. 150 ms → 40 ms with higher read accuracy.
• ·Continuous-learning loop on a 2-week cadence: live misclassifications → retrain → mAP50 gate → hot-swap on Triton with zero downtime.
• ·Vehicle-attribute microservices (color, type) at 0.92+ Macro-F1 with Grafana SLOs and A/B rollout.
• ·Owned T4 fleet scheduling for 10 production models: per-card placement and pod autoscaling to prevent OOM under load.
• ·Agent-driven sweep on input size per model: cut weights to 1.2 MB and inference to 30 ms/frame across all prod models.

Scala · Redis · MySQL · Kafka · AWS · Gurgaon

• ·Personalization API using multi-armed bandits (ε-greedy, Thompson Sampling, UCB) for homepage engagement. Benchmarked variants and shipped the best to production.
• ·Trained an RLHF reward model on annotated user-preference data to rank itinerary recommendations; integrated into the bandit policy for online A/B.
• ·NLP-based resume screening using text similarity and keyword extraction for the Data team's hiring pipeline.

Python · PyTorch · LLMs · RAG · Agents

• ·TA'd a 16-week live AI engineering cohort: ran weekly classes on RAG, agents, fine-tuning, and inference for 200+ engineers. Designed assignments, debugged student projects, took office hours.
• ·Parent platform: InterviewReady. Restructured content for 10k+ subscribers; technical posts reached 20k+ readers.

Python · PyTorch · ML systems

• ·Contributed 300+ implementation problems, ML system-design problems, interactive quizzes, paper-implementation questions, and the study plan for pro users. Platform serves 40k+ users.