Brainstrom A.I.
Agentic AI systems, built with human gates.
I am Steven Novotny. I help people use AI to become more effective than they imagined, by handing them a team they could not have had.
What I build
Agent teams with human gates
Implementer, reviewer, and supervisor agents working over a shared bus. The pass-or-fail bar is written before the build starts, and a reviewer re-runs the work instead of trusting a report. People keep every decision that matters.
Grounded assistants and evaluators
Retrieval over your own material with citations, and multi-agent scoring against explicit rubrics. Every AI score carries a human approve or override step, and users are told when AI produced it.
Serverless delivery on AWS
Lambda, Bedrock (Claude, Nova, Titan), S3 Vectors, API Gateway, and CloudFront, all in Terraform. Small, cheap to run, and reproducible from code.
Recent work
CubeSat workforce training
A live pilot for a statewide university collaboration: self-study grounded in the real course, multi-agent evaluation against versioned rubrics, and an instructor sign-off on every score.
Ministry toolkit
Eight serverless endpoints a church app calls through middleware: scripture and source search, lesson and sermon drafts, and prayer-request triage that flags urgent needs for a human care team.
MoneyPenny
A chief-of-staff layer for Claude Code: one canonical brief of how I work, served into every AI session, with playbooks for adversarial review and honest negative results.
Background
Physicist turned AI engineer
PhD in computational physics and astronomy, an MS in computer science (AI and machine learning) from Georgia Tech, and twelve years teaching physics at the US Air Force Academy. Earlier work on autonomous systems and NASA ISS telemetry. Army cavalry scout, then Air Force officer in space research and development.
Ask my digital twin
It runs on Claude and answers from a brief I wrote about my background and projects. It is an AI, not me.
Chat with the twin