<p>Designed and built "Code Fusion" - a multi-agent platform using AutoGen, Streamlit, Azure OpenAI/OpenAI APIs. Implemented a Gateway agent that routes requests to specialized sub-agents for Python 2-3 conversion, unit test generation, code optimization, and output validation using a planner-like orchestration pattern.</p><p></p><p>Enabled multi-agent collaboration through structured tool calling and handoffs. Built validation agents that return PASS/FAIL verdicts. Implemented auto-retry loops for failed outputs with zero manual intervention.</p><p></p><p>Developed a multi-layer evaluation pipeline: AST parsing for syntax checks followed by validation agents for functional accuracy. Tracked metrics including validation pass rates, response times, token usage, and per-model costs to drive continuous prompt and model improvements.</p><p></p><p>Mitigated failures and hallucinations by moving deterministic tasks like Python 2-3 conversion to tools instead of LLMs. Supported large repositories via per-file processing, intelligent filtering, and chunking. Added extensive logging and analytics for production debugging.</p><p></p><p>Managed state using Streamlit session state for files, outputs, and user selections. Used ConversationSummaryBufferMemory for context retention and ChromaDB with filename/source/timestamp metadata for persistent retrieval.</p><p></p><p>Built ingestion pipelines for ZIP files and GitHub repositories with per-file processing. Integrated logging and token/cost analytics dashboard. Platform deployed in production for code modernization and unit test automation in RAG & Document Intelligence for Medical & Life Sciences.</p><p></p><p>Implemented hybrid retrieval with ChromaDB vector search + BM25 keyword search using MultiQueryRetriever and EnsembleRetriever. Applied FlashRank reranking for precision, outperforming single-retriever baselines on medical document queries.</p><p></p><p>Used LangChain for RAG pipelines, document intelligence, and tool-enabled agents. Leveraged Azure OpenAI, OpenAI APIs, and LlamaIndex for ingestion and structured retrieval. Delivered GenAI 2.0/3.0 layers for clinical document analysis. Integrated GPT-4 Vision for multimodal medical image analysis.</p><p></p><p>Provisioned AWS infrastructure using reusable Terraform modules with Terragrunt DRY methodology across dev/staging/production. Configured EventBridge, SQS, SNS, S3, API Gateway, Lambda with custom authorizer, RDS PostgreSQL with schema-level permissions, and least-privilege IAM roles.</p><p></p><p>Ensured security and compliance by configuring WAF rate limiting for API Gateway, implementing Lambda authorizer for authentication, managing IAM boundaries and PostgreSQL permissions, and enforcing code security with SonarQube and Checkmarx SAST.</p>
