Debanirmalya Patra

Architected a stateless Live Call Transfer routing system using a decorator-based handler registry, enabling self-registering integrations with zero core-logic changes; reduced onboarding time from 23 days to 1-2 hours. Built a multi-modal recovery pipeline (Voice WhatsApp) as an LLM tool call, capturing typed user input via webhooks and rehydrating LLM context; reduced low-confidence failures by 40%. Proposed and implemented a batch transcription pipeline over offline call recordings to recover missed entities from noisy audio, improving data accuracy by 20%+ in downstream verification workflows. Improved STT accuracy by 30%+ using Deepgram Keyterm Prompting (Nova-3), injecting domain-specific vocabulary to increase keyword recall and reduce transcription errors in downstream LLM workflows. Enforced concurrency safety using Redis locks and idempotency controls, reducing duplicate executions by 70%.

Agentic AI Parliamentary Q&A System: Engineered an end-to-end RAG system using LangChain to query a corpus of 5000+ parliamentary records, reducing manual research time for policy analysts by over 40%. Implemented a multi-step retrieval pipeline leveraging query deconstruction, boosting factual accuracy by 30%. Designed a hybrid multi-tool framework using BM25 and FAISS, that dynamically selects from 3+ retrieval methods (filtering) and incorporates a self-correction loop, enhancing retrieval precision by 80%. Deployed using FastAPI, Next.js and AWS, delivering citation-linked answers with 70% faster research turnaround. Improved performance by 25% through rigorous evaluation of faithfulness and answer relevance via RAGAs. ABDM Microsite: Spearheaded the development of a high-performance portal using Django, Next.js and Redux streamlining registration tracking for more than 100k facilities and professionals and obtained Web Application Security Audit clearance. Designed robust RESTful APIs for comprehensive analytics and reporting features, driving 30% increase in adoption, optimized monetary rewards disbursement, and eliminating over 100 man-hours in oversight efforts. Boosted page load performance by 90% through frontend code-splitting with Next.js SSR, while designing role based access control (RBAC) for 4 user levels and conditionally rendering tailored interfaces. Integrated automated Cypress E2E testing into CI/CD pipeline, achieving 80% pre-production defect detection.