Kansas City, Missouri
Jack Stepanek
Software engineer at Veterans United, working on conversational AI for VA lending. On the side, I build open-source tools for evaluating LLM agents.
Now
I’ve been at Veterans United since 2022, most recently as part of the team that took a conversational prequalification agent for VA loan applicants from early scoping through launch. I still work on it every day, supporting new features, debugging, and helping other teams get up to speed as more people build on it.
The parts of the job I keep coming back to are the people parts: working through tradeoffs with product and architecture, serving on our engineering interview committee, and helping teammates untangle hard problems. The projects below grew out of the same instinct: wanting better answers to “how do we know this agent is actually doing a good job?”
Projects
- Waypoint
Trajectory evaluation for tool-using agents. Scores the path an agent took (tool choice, arguments, ordering, efficiency), not just whether the final answer came out right. Ships with a LangGraph compliance-checking agent as the reference system under test.
- Anchor
Claim-level groundedness checking for RAG agents. Splits an answer into individual claims and verifies each one against the context the agent actually retrieved, so a report shows which sentence lacks evidence, not just a faithfulness score.
- ms-pacman-ai
A Rainbow DQN agent that learns Atari games from raw pixels and a score, with no knowledge of rules, enemies, or objectives.
In progress: Tabula, an agent that starts out knowing nothing and learns about the world only through conversation.