Prateek Mehra

Applied ML Engineer building production computer vision and GenAI systems

4+ years experience · Bengaluru

I build and run eight production computer vision pipelines that process 100k+ images a day across 550+ outlets, and cut manual audit effort by 75%.

550+ outlets 100k+ images/day 7,000+ cameras 75% less manual audit

Experience

Applied ML Engineer - Data Tech, EatClub Brands (Box8)

Bengaluru · June 2022 — Present

  • Vision at scale: built eight production pipelines for attendance, OCR, grooming and order checks, from data curation to deployment and threshold tuning.
  • Edge face recognition: built CCTV attendance with liveness checks and an INT8 OpenVINO spoof detector (75 ms/frame), cutting proxy attendance from 22% to 2%.
  • GenAI: shipped a Slack RAG assistant over 16 repos (0.90 Hit@10) used by 20 engineers, and a VLM-based audit that catches faulty CCTV cameras.
  • MLOps: moved 15 model repos onto one shared GitLab CI pipeline and cut Docker images from 182 GB to 26 GB (86% smaller).
  • Mentored five junior ML engineers and planned the team's quarterly roadmaps.

Currently exploring: edge AI · real-time CV · multimodal systems · low-latency inference

Blogs

When Are You Done Speaking? Frozen Whisper Embeddings vs Handcrafted Features

Turn detection with Whisper Tiny + XGBoost: an ablation study, a scaling curve, and one tuned threshold
Published · Updated · 5 min read

On-Device Vision at 4% CPU: FP16 Core ML vs PyTorch MPS

YOLOv8n bottle detection on an M1 Air that sleeps between frames
Published · Updated · 5 min read

Skills

Computer vision: YOLOv8, OpenCV, face recognition, OCR, OpenVINO, Core ML
GenAI: RAG (pgvector, BM25), VLMs, Whisper ASR, LLM evaluation
ML: PyTorch, TensorFlow, XGBoost, scikit-learn
MLOps & infra: Airflow, Docker, GitLab CI, AWS (EKS), GCP (Cloud Run)
Languages: Python, SQL