At Lab2Market Rail (IISc Startup), I trained and fine-tuned YOLO object detection models and developed ResNet-based image classification and PaDiM anomaly detection pipelines for railway undercarriage defect inspection as part of the MVIS project.
I also implemented Python data-processing and inference pipelines for the railway inspection system. I analyzed model predictions and inspection alerts, investigated false positives, and recommended out-of-distribution validation after identifying potential train-validation data leakage.
I built AgentGate, a prototype policy gate for AI agent tool calls. It checks role identity, argument safety, request intent, and call rate, and can allow, deny, or hold an action for approval; held actions are re-checked before execution and decisions are recorded in an audit file.
I connected AgentGate to a LangGraph workflow and tested it against prompt injection using a local 3B model; the gate blocked all four attempts. I also built a YouTube Q&A agent that retrieves transcript content using embeddings and semantic search to answer questions with context.
