I build AI agents that answer from real data, and the pipelines underneath them.

Built with Langflow and FastAPI. A router sends each question down a semantic or an analytical path, and eight live J&J APIs ground the answer in that customer's real-time data.

I deployed the custom backend and the Langflow workflow on AWS Bedrock AgentCore myself, with OAuth 2.0 token-based authorization verified against the client's database and user-scoped conversation storage in RDS PostgreSQL.

I built the Python, FastAPI, and Azure OpenAI backend. The specialist agents share a base class, a registry, and one API, so adding an agent takes one implementation file and one registry entry.

Outputs are grounded in the source input, which removed invented names, dates, and figures. Test coverage is calculated from story-to-test mappings. Reviewers get section-level edit suggestions, open-question resolution, and recorded sign-off before export.

The pipeline does multi-step reasoning over Pinecone vector search, OpenAI embeddings, and knowledge graphs. Python and Spark pipelines feed it by unifying structured and unstructured sources across 8 financial use cases, with automated data quality checks.

I designed the evaluation framework that gates releases: LLM-as-judge scoring, canary deployments, and audit-driven feedback loops. I also shipped 7 production NLP and LLM models, fine-tuned with LoRA and deployed through Airflow CI/CD with PII guardrails.