Projects

RAG evaluation platform

Developed an evaluation platform that logs generated answers, retrieved context, and failure cases from a RAG pipeline. A correction loop increases retrieval depth and regenerates answers that fall below a quality score of 70.

FastAPI, Streamlit, ChromaDB

How it works

The platform tracks faithfulness, answer relevancy, and context precision across runs. Each answer is limited to one correction attempt, with both outputs retained for comparison.

  1. Retrieve context and generate an answer.
  2. Evaluate answer quality and log the result.
  3. Increase retrieval depth and regenerate answers scoring below 70.

EV anomaly detection

Co-developed an anomaly detection framework for electric two-wheelers using a GRU-based autoencoder. The model uses over 100,000 telemetry records to identify abnormal driving patterns and system-level anomalies through unsupervised sequence modelling.

GRU autoencoder, unsupervised learning, time-series data

Patent application

A Device and Method for Detecting Anomalies in Electric Two-Wheelers.

Published application no. 202541080133.

With Dr. Ashok B, Dr. Gunavathi C, and Rudraksh Srivastava.

Additional projects

LLM-based data auditor

Developed an auditing pipeline that combines retrieval-augmented generation with structured validation to identify inconsistencies in datasets and documents.

Voice-to-data assistant

Built a real-time voice interface that converts unstructured spoken commands into structured records for downstream data processing.