Enterprise AI, from the orchestration layer down

Internal product. Contributed across R&D, design and implementation of the platform's agent orchestration, tooling and evaluation layers as part of the engineering team.

• Developed an LLM gateway integrating OpenAI, IBM watsonx, and Meta Llama models through Azure and direct provider APIs, enabling provider-agnostic model selection and avoiding single-vendor lock-in.
• Designed MCP servers using SSE and stdio transports to standardize tool access across enterprise agents.
• Implemented multi-layer agent memory workflows — short-term, long-term and episodic — for contextual continuity across sessions.
• Built an LLM-as-judge evaluation harness that scores agent responses against a labeled query/response dataset.
• Implemented input/output safety guardrails covering PII/PHI, jailbreaks, harm, social bias, profanity, and prompt injection, alongside IBM Guardrails.
• Integrated external data and tool nodes spanning social platforms, medical sources (PubMed, DrugBank, ClinicalTrials.gov, FDA), and relational databases (PostgreSQL, MySQL).
• Delivered a social-sentiment analytics workflow orchestrating agents to pull and analyze posts and comments across multiple platforms.

Research-facing Q&A over clinical trial, drug and disease data. Started as a traditional RAG pipeline and was redesigned as a Graph RAG once relationship modeling became the bottleneck.

• Built and evaluated metadata-aware RAG using hybrid semantic and BM25 retrieval over clinical trial, drug, and disease data; identified its limits in modeling many-to-many drug–disease–dosage relationships.
• Contributed to the Graph RAG redesign using Neo4j to model drug–disease–dosage relationships, and evaluated AWS Neptune for larger-scale ingestion.
• Developed natural-language-to-Cypher and natural-language-to-Gremlin query translation workflows for graph-based clinical question answering.
• Iterated on embedding and metadata strategies to improve entity-relationship linking, informing the team's subsequent RAG architecture decisions.

A multi-stage pipeline that ingests raw, multi-source data and restructures it into tagged, categorized records for downstream querying.

• Developed 10+ specialized LangGraph agents for sentiment analysis, thematic clustering, and entity extraction within an orchestrator/sub-agent architecture.
• Built MCP servers exposing schema-introspection and query-execution tools for natural-language-to-SQL workflows.
• Contributed to multi-stage data structuring pipelines using topic modeling to group related content before records are written to the database.
• Embedded business-logic and governance rules as structured context files to keep agent-generated answers aligned with client data-governance policy.

Agent-driven conversion of enterprise Tableau dashboards into Power BI equivalents, reducing manual rebuild effort.

• Developed LangGraph/LangChain agents for Tableau-to-Power BI migration, achieving 85–90% conversion accuracy under project evaluation criteria.
• Built custom parsers for Tableau (.twb) XML files to extract chart types, calculated fields, filters, and axis configurations.
• Used LLM agents to auto-generate equivalent DAX queries and Power Query (M) scripts for Power BI.
• Developed the FastAPI service layer exposing the chatbot and automation agents to the rest of the migration pipeline.