jaswanth mada

About

Built and productionized an internal LLM-powered document Q&A system at Morgan Stanley using a LangChain-based RAG pipeline (FAISS + OpenAI) with AWS ingestion (S3/Lambda), handling 100k+ pages and cutting lookup time ~35% while keeping responses under 3 seconds. Strong on reliability: automated evals/CI (pytest + GitHub Actions), CloudWatch monitoring, drift detection (prompt drift and fraud-model drift), and security controls (IAM + app-level authorization) in a financial-services environment.

Experience

Education

Key Strengths

• Built and deployed internal LLM document Q&A/RAG system at Morgan Stanley (100,000+ pages ingested)
• Reduced end-user document lookup time by ~35%
• Improved grounded/citable responses; reduced partially grounded responses by ~25% via chunking + retrieval depth + prompt/citation tuning
• Designed structure-aware chunking strategies (token size/overlap tuning; markdown structural splitting) to improve retrieval consistency
• Implemented secure access control in RAG pipeline using AWS IAM plus application-level authorization
• Maintained sub-3-second latency while managing inference cost through CloudWatch instrumentation and tuning
• Established automated evaluation and CI gating (pytest + GitHub Actions) for prompt/retrieval changes before deployment
• Proactively detected and mitigated prompt drift after OpenAI API updates using versioned prompts and automated response-structure tests
• Detected model drift in fraud classification using KS and precision-recall monitoring and triggered retraining before production impact
• Effective collaboration with non-technical stakeholders (research/compliance) to shape requirements and validate output quality