Computer Vision Engineer for Live Analytics

Sahil Yadav

LinkedIn +91-706******* Email

EDUCATION

IIT Roorkee Roorkee, UK, India

WORK EXPERIENCE

Computer Vision Engineer YouInSports Pune, India Mar 2025 – Present Real-time video analytics for chess and basketball live streams: 2 AI monetisation features lifting engagement by 12–15%.

• Deployed two end-to-end live video inference pipelines on GCP Compute Engine VMs and Cloud Run, scaling past 500 live streams a day with Firestore backing event, state and per-stream telemetry storage for both products.

• Chose MobileNet CNN backbones for piece and hand detection to hold a real-time latency budget on commodity compute, training on 100K synthetic + 10K real samples to reach production-grade accuracy on live streams.

• Cut per-frame compute cost by rectifying all 64 board squares once via RANSAC homography, then running Hungarian matching across subsequent frames to emit PGN move commentary without re-detecting the board.

• Ran transformer-based vision models in production – RF-DETR detection, SAM 3.0 segmentation and SmolVLM jersey OCR at 92% – to rebuild basketball possessions and games in 2D from ordinary broadcast video footage.

Research Engineer, AI Evaluation/Benchmarking Mercor Remote Mar 2024 – Apr 2025 Frontier-model evaluation programs: eval design and expert-annotated data across 4K+ complex multi-turn agentic tasks.

• Owned and built the data layer for model training and evaluation end-to-end – task definitions, annotation schemas and grading rubrics – lifting inter-annotator agreement to 0.87 Cohen’s κ across 4K+ agentic evaluation tasks.

• Acted as the quality gate on every incoming data batch, cutting label noise by 22% before researcher handoff and converting 300+ agent failure edge cases into reusable new benchmark sets for later evaluation and regression rounds.

• Partnered closely with research and infrastructure teams on reward checks and feedback loops across RL environments, lifting agentic benchmark pass rate by 14% and root-causing failures across long multi-turn agent trajectories.

Computer Vision Engineer Professor Siddhartha Khare Dec 2023 – Sep 2024 Self-driving car perception stack trained on approx. 40–42K Canadian driving images, with 9–10K held out for testing.

• Parallelized training across multiple GPU nodes on an HPC cluster using PyTorch DistributedDataParallel, cutting end-to-end training time on a 40–42K image driving dataset and freeing cluster hours for further experiments.

• Benchmarked YOLOv9 against YOLOv8 on identical hardware, reducing inference latency by 8% and raising detection accuracy by 7% to a final 85% on a 9–10K image test split, then promoted the faster model to production.

• Built and maintained the supporting data pipeline – annotation in LabelImg and Roboflow, augmentation, class balancing and train/test splits – to keep every GPU node fed throughout long multi-node distributed training runs.

Machine Learning Engineer Professor Sharad Jain Feb 2024 – Apr 2024

• Designed LSTM rainfall forecasting models over 15K samples of river-flow and meteorological variables, improving accuracy by 18% over Regression, Random Forest, AdaBoost, Gradient Boost, XGBoost and RNN baselines.

PROJECTS

Text-to-Image Generation System SocBiz, IIT Roorkee

• Formulated a Transformer-based encoder with StyleGAN to generate high-resolution images from text prompts.

• Established end-to-end preprocessing, training and validation pipelines over a large-scale dataset ( 1GB), refining semantic comprehension by 25% through targeted fine-tuning of the text encoder on domain-specific captions.

RAG-Based LLM Application on Real Estate Geographical Data Data Science Group, IIT Roorkee

• Built a retrieval-augmented serving path with LangChain-Pathway, improving query latency by 15% and retrieval accuracy by 32% over the baseline retriever using zero-shot inference, prompt templating and entity disambiguation. Fine-Tuned LLM Architecture for Retrieval-Augmented Generation and Decision-Making Self-project

• Engineered a high-performing retrieval-augmented LLM application through advanced prompt engineering and systematic hyperparameter tuning across both the retrieval and the generation stages of the serving pipeline.

• Developed a high-precision fine-tuned LLM (0.7–0.8) for domain-specific question answering and downstream decision-making, achieving a 20% increase in contextual response accuracy over the untuned base model.

POSITIONS OF RESPONSIBILITY AND EXTRACURRICULAR ACTIVITIES

Himalayan Explorer’s Club Executive Member IIT Roorkee, May 2023 – Present

• Organized an adventure expedition for over 200 participants, coordinating safety briefings and ground logistics. IIT Roorkee Taekwondo Club Member IIT Roorkee, Feb 2023 – Present

• Engaged in over 15 competitive events, tournaments and team-building activities across two years of regular training.

Contact this candidate