I build AI development tooling, and I've spent 20+ years helping financial institutions decide which intelligent systems are worth betting on.

By day: Manager, Financial Systems at American Express — product owner for GL reconciliation and the data layer underneath finance operations. Data governance at a scale where "mostly right" isn't an option.

By night: Building in the agentic-coding space. ralph-claude-code grew to ~9.7k stars and 700+ forks, and running it in the wild taught me where AI development is actually heading: the loop is getting commoditized. The hard, valuable problems are now upstream (what are we building, and how will we know it's right?) and downstream (can we prove it works before it ships?).

In a past life: Co-founder and executive at fintech startups, and one of the creators of bank pricing-optimization technology (Nomis Solutions, SAP/Khimetrics, Earnix, Response Analytics). Industry analyst at Tower Group (now Gartner), telling banks which analytics/ML technology actually mattered and telling vendors what banks actually needed. Director/VP roles at US Bank and Truckstop.com. Author of Seven Billion Banks on personalized banking.

AI agents write the code. CodeFRAME owns everything before and after.

• Think — guided requirements discovery → PRD → dependency-aware task breakdown
• Build — hands tasks to the coding agent you already use (Claude Code, Codex, OpenCode, Kilocode, or a built-in agent), with stall detection, retries, and batch runs
• Prove — PROOF9, nine verification gates where every captured failure becomes a permanent requirement future builds must satisfy
• Ship — cf pr merge only merges when proof passes

My working thesis: code quality management starts with good requirements. You can't prove something is right if you never wrote down what "right" means.

I write about this at Code & Context — AI-assisted development, proof over review, and building production-ready systems.

pip install codeframe-ai

The project that started it all: continuous autonomous development cycles with intelligent exit detection, rate limiting, and circuit breakers. With long-running goals now native in Claude Code, the loop itself is largely solved. The open question is the one Ralph users kept running into: how do you know what the agent shipped is actually right? That's where my attention has moved — PROOF9. Start with PROOF9: Review the Statement, Not the Proof.

Upload a spreadsheet, walk through eight guided stages (profiling → cleaning → feature selection → AutoML → evaluation → predictions → API deployment), with plain-language explanations at every step. Next.js + FastAPI.

• Proof-native forge — what would GitHub look like if PRs were built around proof obligations instead of code diffs, and agents worked in the fast lane alongside humans?
• Formal math knowledge layer — a declaration-centric contribution layer over Lean mathematics.

I'm a mathematician at heart (MS Mathematics, MEng in Mechanical Engineering (Systems Engineering); algebraic topology and K-theory for fun) and teach college math as an adjunct instructor. Six Sigma Black Belt, MIT COO certificate.

I'm also a composer for film and media, a photographer, and I speak German fluently plus enough Russian, French, and Italian to get into trouble.

I hosted The 6 to 7 Figures Show, a podcast on scaling consulting practices — the archive is still up, and something new is in the works.

I do advisory work on AI strategy, agentic development, and technical due diligence through BSG (Bria Strategy Group).

• 🌐 frankbria.com
• ✍️ lab.frankbria.com — Code & Context blog
• 🔗 LinkedIn
• 📬 frank@frankbria.com

"I build intelligent systems — and help people decide which ones to bet on."