Since February 2026 I have used AI coding agents to build software end to end: a governance and routing layer for AI agents, a control plane that connects assistants to infrastructure, a learned routing model, iPhone apps that are live on the App Store, and sites built with and for the people in my life. AI agents write most of the code; I set the specification and the bar it has to clear.
These are personal projects, built outside work and independent of Acronis. My day job is partnerships; this is how I stay close to what AI platforms can actually do.
AI systems
Three systems that work together: one governs and routes what agents do, one gives assistants a safe way to act on infrastructure, and one learns how to spend less on the work.
Feb 2026–present
BrainstormRouter
An authorization, evidence and routing layer for AI agents. Every agent action runs under a cryptographically signed, scoped capability grant, and every decision, allowed or denied, lands in a tamper-evident ledger. Inside that boundary the router picks the best model for each turn.
Ed25519-signed capability grants with verifiable delegation: an agent can pass a narrower slice of its authority to another agent, never a wider one.
Hash-chained, append-only audit ledger, so denials are evidence too.
Thompson-sampling model routing across 55 curated models from nine providers, with real-time budgets and memory refresh.
Drop-in OpenAI- and Anthropic-compatible API, plus a Model Context Protocol (MCP) server.
Authority is checked before any model is called, and the ledger records denials as well as completed calls.
A governed control plane that connects AI coding assistants, such as Claude Code, to infrastructure through one channel. The question it answers: how do you give an assistant useful access to real systems while keeping its actions understandable and controllable?
MCP server with 58+ tools behind a single platform contract.
Approval workflows, cost tracking and audit reporting for agent actions.
Evaluation tooling for model output, a command-line interface and a desktop app.
A small research project with a hard rule: ship only what clears a bar set in advance. It tries to predict which phases of a software-development pipeline a task actually needs, so an agent can skip the rest, trained on 2,203 real Claude Code session turns.
Gates set before training: a mean F1 of at least 0.75 under five-fold cross-validation, and at least 25% lower cost with no more than a 5% quality drop against running every phase.
A one-shot classifier scored 0.52 and failed. A sequential version that conditions each phase on the ones before it reached 0.61: better, still short of 0.75, so it stayed a research result rather than a routing decision.
The quality gate needs build-and-test outcomes from full pipeline runs, which the session data does not have; that is the next dataset to collect.
A separate model-tier classifier trained on 400,000+ RouterBench data points reached 93% cross-validated accuracy.
Native iPhone and web apps, built with the same agent workflow and taken through to real users.
Mar 2026–present Built for Rebecca
Our Book Nook
A private reading journal, built for Rebecca, that a book club can share. Write a thought the moment it lands or say it out loud, bring in Kindle highlights with one tap, and keep it all private until you hand a book to a friend or open it with your club. Even then, nothing shows before they have read that far.
A college football site Cooper and I are building together: scores, polls, team histories and headlines in one place, next to Cooper’s own takes. Cooper chose the design and writes the takes; I built the platform and the data behind it.
A DuckDB warehouse of every team, game, player and season, fed from CollegeFootballData.com behind a call-budget guard, that publishes page-ready data to the site.
Elo ratings built only from earlier games, which give each matchup a win probability on the home page.
Runs on Cloudflare Workers with D1 and R2, on the EmDash content system.
A record of growing up through youth sports: a child's seasons, coach notes a parent can read and growth a kid can see, in one record that compounds year over year. League operations run simulation-first: intent, simulated change, human approval, execution and undo, the same agent-safety pattern as my other systems. Web and SwiftUI iPhone clients share one platform.
The common thread is a disciplined way of working with agents: they write most of the code, and I set the bar it has to clear.
Spec first
Substantial work starts as a written specification with acceptance criteria, so an agent's output is judged against something other than its own description.
Gates, not vibes
Results have to clear a bar set in advance. BrainstormLLM's phase predictor improved from 0.52 to 0.61 F1 and still fell short of its 0.75 gate, so it did not ship.
Review loops
Independent reviewer agents score the work against a stated bar, round after round, and a change lands only when the findings are fixed or explained.
Verify on the real path
Done means proven where it runs: a live response, a database row, an App Store build, not a passing test in isolation.