At Cognizant Technology Solutions, I re-architected a legal document recommendation workflow into a LangGraph-based Agentic AI system. It processes 100+ alerts a day, with approximately 30% routed through the agentic path.
I designed a practice-area-aware alert triage classifier using Claude Sonnet, achieving 98% recall on a labeled benchmark of 10,455 alerts. For document recommendations, Claude Sonnet orchestrates retrieval and Claude Opus evaluates results, reaching approximately 92% recall and 93% end-to-end accuracy across three jurisdictions.
I also built a RAG microservice to augment GPT-4o for document-based clinical translation, achieving 94% accuracy, and Azure and Whisper Large V3 Turbo microservices for live and multi-speaker audio translation. I replaced 14+ hours per week of manual data extraction with a Looker Studio sales dashboard.
In my Self-Correcting Content Pipeline project, I combined CrewAI and LangGraph to research, draft, critique, and revise articles. I added guardrails, testing, telemetry, and GitHub Actions CI/CD; across eight runs, the mean judge score rose from 6.6 to 8.1 at under $0.06 per article.
