rishabh jain

I’m an AI/ML Engineer building production-grade computer vision and LLM systems. I focus on fine-tuning multimodal models (YOLO, VLMs), designing RAG and LLM evaluation pipelines, and deploying scalable ML microservices.

At Coschool AI, I fine-tuned YOLOv11 (nano & small) for worksheet label/separator detection and served it via Kubernetes-hosted FastAPI. I built an end-to-end worksheet segmentation pipeline that combines YOLO detection, VLM-based matching, cross-page stitching, and missing-separator recovery—processing 10k+ handwritten submissions with parallelized S3 upload and inference.

I also developed multimodal OCR workflows using Mathpix + GPT-4o vision + Claude, with image quality validation, diagram extraction, and LaTeX formatting preservation for complex handwriting. To improve efficiency, I replaced per-request LLM off-topic detection with embedding + GMM classification using OpenAI embeddings and Milvus, cutting inference cost by ~50% at comparable accuracy.

Beyond modeling, I lead production reliability and iteration speed: I built an LLM evaluation framework benchmarking GPT-4o-mini vs GPT-4.1-mini, migrated 115 prompts to Langfuse, and implemented a production Graph RAG pipeline for stronger factual grounding. I’ve also developed agent-based NL2SQL/GraphQL systems powering dynamic analytics dashboards for cost, usage, and quality metrics.