Projects

Multimodal GenAI Compliance System

• Architectural design images were reviewed manually against internal policy standards, slow, inconsistent, and difficult to scale across a large regulatory body.
• Built a multimodal GenAI pipeline where users upload design images and vision agents automatically check them against policy rules. Output is a structured compliance report across three categories: Compliant, Non-Compliant with detailed reasoning, and Needs Review.
• Integrated MinerU, a non-OpenAI-compatible model, by building a custom FastAPI wrapper and deploying it on the Katonic platform. Full agent architecture built using Groq and custom-deployed models.
• End-to-end ownership across architecture, development, customer stakeholder management, and internal product alignment.

GenAI-Enabled QA Automation Engine

• A 60-member QA team spent significant time and cost on testing despite heavy automation investment. The end-to-end QA lifecycle was still largely manual.
• Built a GenAI engine that automated the full QA lifecycle, from user story to test case generation, and from test cases to executable test scripts. Removed the most repetitive steps from the process entirely.
• 33% productivity gains across end-to-end QA test cycle execution, freeing the team for higher-value development work.

Private LLM-Powered Customer Support Chatbot

• Customer support queries required searching across large volumes of technical documentation, with resolution times running into days.
• Built a conversational AI system using Advanced RAG, with open-source LLMs identified and deployed privately on AWS EC2 to meet strict data security requirements.
• Query resolution time reduced from days to minutes. Response accuracy exceeded 95% in production.
• Also contributed NLP pipelines for named-entity extraction, Flask REST APIs for full-stack integration, and database design across the solution.

Text-to-SQL for Leadership Analytics

• Leadership depended on analysts to pull reports from complex databases, creating delays and a constant bottleneck on the data team.
• Built a Text-to-SQL system using a RAG pipeline where users ask in natural language and get SQL-backed results with automated dashboard output. A Schema Intelligence Agent maps natural language intent to the right tables, columns, and relationships. A Dashboard Generator Agent then converts query results into visual reports for leadership.
• Reduced dashboard access time and removed analyst dependency for routine reporting.

AI-Driven Mental Health Platform

• Access to structured, consistent mental health support is limited, especially in schools and clinical settings where professional bandwidth is a real constraint.
• Built a CBT-based platform providing AI-driven therapy conversations and patient assessments. Deployed and used across schools in North India and piloted at a major hospital in Gurugram.
• Led product and UI development using React JS, working directly with clinical practitioners to align the experience with therapeutic frameworks.

D2C E-Commerce Intelligence Platform

• D2C brands had no reliable way to track competitor pricing and market trends at scale without heavy manual effort.
• Built a scalable backend for price tracking, historical analysis, and trend prediction across multiple e-commerce platforms.
• Core component is a multi-agent web scraper for large-scale data collection, handling dynamic content, rate limiting, and data validation automatically. Reduced data processing time by over 60%.

POC Lab

A selection from a wider body of experimentation spanning telecom, enterprise software modernization, and multi-agent AI systems.

• Log Lifecycle Automation (Telecom). Automated the full lifecycle of log management for customer support engineers at a major telecom client. Goal was to reduce resolution time from days to under 2 hours by intelligently triaging and routing log events.
• Java to Microservices Migration (CrewAI). Agentic solution that analyzes legacy Java monolith codebases and generates microservice architecture recommendations, reducing manual effort in large-scale modernization work.
• Multi-Agent E-Commerce Scraper. Adaptive scraping framework with domain-specific agents capable of handling a wide variety of site structures reliably at scale.
• RAG Pipelines with Multi-Agent Orchestration. Multiple RAG implementations across domains, with orchestration layers managing retrieval, reranking, and response synthesis across specialized agents.