Headline

AI Engineer building GenAI & LLM Systems | Production ML | Publicis Sapient

About

AI Engineer focused on building and scaling GenAI systems in production. At Publicis Sapient, I work on Bodhi Atomize — a multimodal AI platform that transforms images, videos, and GIFs into structured signals such as text, characters, emotions, and branding elements using large language models and computer vision. The objective is to reduce manual effort and enable reliable enterprise-scale content understanding. My work sits at the intersection of GenAI systems, computer vision, and production ML engineering. I design and operate pipelines that handle real traffic, where concerns like latency, token limits, retries, backpressure, and system failures matter as much as model quality. I’ve built production workflows using Gemini 2.5 Pro, implemented structured outputs with Pydantic, and worked on evaluation frameworks using DeepEval. A significant part of my learning comes from running systems under real-world constraints rather than isolated experimentation. Previously, I worked on object detection and defect detection systems, improving both accuracy and inference performance across large image datasets. What I care about: • Building AI systems that reliably work in production • Understanding failure modes early and designing for robustness • Writing clean, scalable ML and backend code • Learning from real usage patterns, not just research papers Open to conversations around GenAI systems, LLM infrastructure, ML engineering, and production AI challenges.

Location

Bengaluru

AI stack

claude code

claude

gemini

hugging face

langgraph

vLLM

Work experience

Senior Associate Data Science L1 at Publicis Sapient, Jun 2025 - Present · 1 yr 3 mos, Bengaluru, Karnataka, India, - Worked on Bodhi Atomize, a multimodal GenAI product for enterprise marketing analytics, decomposing 10,000+ image, video, and GIF assets into structured JSON outputs for large clients like Eli Lilly. - Built and productionized LLM pipelines using Gemini 2.5 Pro, handling multiple parallel calls per asset with structured outputs using Pydantic - Designed systems to handle high-concurrency and burst traffic, implementing retry strategies with exponential backoff, token budgeting, and backpressure to manage rate-limit failures - Achieved ~95% reduction in manual analysis time (from hours to ~2 minutes per asset) Integrated computer vision + LLM workflows using YOLO, PaddleOCR, and LLM reasoning to extract 50+ structured components per asset - Used DeepEval to evaluate correctness and consistency of LLM outputs, improving reliability for production use - Built and deployed FastAPI-based microservices, working with Redis, Celery, and Kubernetes (KEDA autoscaling) to support 1,000+ concurrent requests with low latency

Data Science Trainee at Lincode India, Apr 2025 - Jun 2025 · 3 mos

Data Science Trainee at Lincode Vision Labs, Oct 2024 - May 2025 · 8 mos, • Integrated RF-DETR into production CV pipelines, achieving ~1.8× faster inference and +7% mAP@50 over YOLOv8. • Improved defect-detection mAP@50 from 55% to 82% using synthetic data and multi-stage training. • Improved client model accuracy by 16% through dataset curation and preprocessing across 10,000+ industrial images.

Data Science Intern at Lincode India, Oct 2024 - Mar 2025 · 6 mos

Junior Machine Learning Engineer at Omdena, May 2024 - Aug 2024 · 4 mos, Remote, • Led the supervised modelling team in predicting urban farming zones in Milan using geospatial data. • Engineered an XGBoost model with 93.68% accuracy for urban farming predictions. • Conducted EDA on 106,000 rows of geospatial data using the Geopandas library to identify suitable farming areas. • Implemented real-time predictions for urban farming zones, optimizing data handling and model efficiency.

Domain Lead(Computer Vision) at Epoch, Jan 2024 - May 2024 · 5 mos, Sri City

Education

Bachelor of Technology - BTech, Computer Science, Indian Institute of Information Technology, SriCity

12th Grade, Dr KKR's Happy Valley School

Public links

github: https://github.com/shivangsingh26