About

CV infra startup · PyTorch · YOLO · PARSeq · Triton · TensorRT · FastAPI · Kubernetes · Docker · Bangalore

• ·Designed and deployed an end-to-end vision system processing live RTSP streams across 10+ detection modules (LPR, phone, helmet, wrong-way, PUCC) at 0.9+ F1 on Indian CCTV infrastructure.
• ·Cut inference latency by 54% to ≤10 ms on edge by replacing anchor-based detection with NMS-free YOLOv10 and tuning Triton / TensorRT serving. Sustained 0.90+ mAP across 8 classes.
• ·Rebuilt the license-plate pipeline: swapped the CRAFT text detector for YOLOv11s and the BiLSTM recognizer for a PARSeq ViT, then exported to TensorRT FP16 for edge. Dropped latency from 150 ms to 40 ms (~100 MB GPU per 100 images) while raising read accuracy.
• ·Built a PUCC pollution-certificate module as a FastAPI service: YOLO confirms the emission probe is inserted in the vehicle, LPR reads the plate, and the detected testing-center name is validated against the government registry of authorized centers.
• ·Built a continuous-learning pipeline that ingests live misclassifications on a 2-week cadence, retrains, validates via mAP50 quality gate, and hot-swaps on Triton with zero downtime.
• ·Shipped vehicle-attribute microservices (color, type) at 0.92+ Macro-F1 with Grafana monitoring, structured logging, and A/B testing for rollouts.
• ·Owned managed deployment of 10 real-time models across T4 GPUs: scheduling which model runs on which card and autoscaling pods to prevent OOM and crashes under load.
• ·Automated agent-driven research to find the optimal image size per model, cutting weights to 1.2 MB and inference to 30 ms/frame across all production 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.
• ·NLP-based resume screening using text similarity and keyword extraction for Data team hiring.

Python · PyTorch · LLMs

• ·Designed and taught an AI engineering curriculum for 200+ engineers in a live cohort. Restructured content for 10k+ subscribers; technical posts reached 20k+ readers.

Python · PyTorch · Markdown

• ·200+ implementation problems on gradients, optimizers, attention, and ML systems. Paired with the AI engineering cohort so each module has matching problems.