Ashish Soni

I've built production-oriented agentic systems, multi-modal RAG pipelines, and real-time computer vision applications as a B.Tech AI/ML student and Back-End AI Engineering Intern at FlyRank Corp.

For ML4Sci DeepLense, I built a multi-class CNN classifier that achieved 92.8% validation accuracy and per-class AUC above 0.97 on a 30k-image dataset. I also designed a Pydantic-validated agentic workflow for the DeepLenseSim simulation pipeline with natural-language parameter extraction, human-in-the-loop clarification, and graceful failure reporting.

I'm an active open-source contributor to TensorMap, where I shipped four merged PRs and fixed a training failure that unblocked the sample pipeline. Across projects including MedGuard, an AI Resume Screener, an AI Research Agent, and a PDF Chat Assistant, I use LangGraph, FastAPI, Docker, and end-to-end MLOps to turn ML workflows into usable systems.