Agasthya Omkumar

AI engineer for Urban and Housing Public Safety project; designed large-scale visual pollution detection workflow and architecture, optimizing automated urban safety inspections. Maintained the world's largest public safety inspection system, processing 6M–8M images/month and detecting 147+ urban safety and housing violations using production-grade ML pipelines. Designed business logic libraries to enforce violation-specific rules, protect PII, and minimize false positives; leveraged LLM-assisted rule interpretation and validation for faster iteration across cities and regulations. Developed violation geolocation estimation using lat/long, camera intrinsics, and GIS integration to support city-level urban planning and AI-generated compliance insights. Integrated depth models to filter distant objects, significantly reducing false positives in open-world detection. Curated training data with advanced techniques (negative data mining, synthetic data generation using GANs and generative augmentation), improving mean precision & recall across all classes by 47%. Built and automated Label Studio pipelines for pre-annotation and auto-annotation, reducing dataset curation time by 64%; integrated Segment Anything Model (SAM2) and GenAI-assisted mask refinement for rapid POCs. Maintained MLOps stack: Triton Inference Servers, Docker containers; enabled scalable deployment of both vision and multimodal (VLM) models in production. Worked on evaluation and integration of Vision-Language Models (VLMs) and LLM-based reasoning layers to support explainability, violation summarization, and next-gen analytics. Oversaw and optimized workflows of multiple annotation teams, ensuring rapid, high-quality dataset readiness aligned with GenAI-accelerated data pipelines. Delivered cross-vertical demos (defect detection in OEM pre-market segment) and led pilots for multiple cities in India and abroad. Contributed to research and integration of SOTA GenAI and multimodal models including Qwen, GLM, and advanced Vision-Language Models (VLMs) for next-gen system enhancements.

Designed and maintained MySQL databases, implementing efficient schemas and relationships. Developed and optimized database queries to support web and enterprise applications. Gained hands-on experience in full-stack development fundamentals and database-driven application design.