
Division of IT at UMD(opens in a new tab)
AI Software Engineer
Part-timeOct 2025 - Present1 yr
- Bot configs went live untested, so I built RAGAS evaluations with auto-generated Q&A sets per bot, raising faithfulness from 0.83 to 0.91 (9.64% improvement).
- Vector search missed prerequisite chains, so I built spaCy NER and REBEL relation extraction into a Neo4j Graph RAG, improving multi-hop accuracy by 2.1%.
- Chatbot evaluation runs were hard to compare, so I built a Vue 3 dashboard with config diffs, cutting review time from 10 mins to 40 secs (94% developer time saved).
- Manual evals didn’t scale, so I orchestrated an async SNS/SQS Lambda pipeline that runs question generation, RAG answering, scoring and storage in one click.
- Bug triage was manual, so I automated bug detection and fixes using Playwright in CI/CD, and on each commit Claude creates fix merge requests for human approval.
- The chatbot had no federal content, so I built Federal Register and GraphQL ingestion connectors for pgvector retrieval, backed by 91+ tests.
- Created a Python & Remotion pipeline to create product demos and walkthrough videos through a claude skill and rendering video's using code.
Tech stack
- typescript
- python
- langchain
- aws
- postgres
- graphql









