Builds
Software, systems, experiments, books, and AI skills. Different forms, one recurring instinct: make the complicated thing usable.
Operator proof
Book + AI skill
A full book and executable AI-agent skill for moving customer support from chatbot theater to a trustworthy operating model.
I wrote the field guide for the AI support migration — and shipped it as a working AI skill.
Voice agent for CS
Voice signals for Customer Success that ring the CS owner, not the customer — the escalation path turned into an agent.
AI pointed at the customer-facing org itself, not at the customer.
Grounded AI agent
A live onboarding intelligence agent that maps how a team actually works from GitLab and Notion signals, then proposes a safe first task.
AI applied to a real org problem — grounded in how teams actually work, not how the wiki says they do.
Competed in the national Google Cloud Rapid Agent Hackathon.
Support CLI
A Go CLI for Intercom and Fin — test AI responses, triage urgent tickets, and read conversation stats without leaving the terminal.
Tooling built for the people actually running the support queue.
Hiring signal
An AI job-fit analysis tool for customer-facing technical roles — built from a career of hiring for them.
I know this talent market from both sides of the table — well enough to ship a product for it.
Crash-safe support agent
A harness for support agents that issue credits, send emails and close tickets. After a crash it asks the provider for a receipt instead of retrying, and hands off to a human when it can't prove what happened.
Agents that act on customers need receipts, not memories.
Evidence check for CS
Checks what you promised a customer against what they can actually use: built, enabled for their account and verified by them are three separate facts. Models read the evidence, fixed rules decide the verdict, and a human approves the update. Nothing sends on its own.
A closed ticket isn't a kept promise — the check to run before the "it's live" email.
Built for the Nebius × NVIDIA Global AI Hackathon.
The Lab
Built because I couldn't not build them. The instinct is the same one I bring to the work.
Political simulation
A 48-turn political simulation that makes sustainable peace winnable, difficult, and legible through the systems surrounding it.
Built at an AWS Loop Hacking hackathon.
Therapeutic software
A therapeutic writing environment inspired by EMDR, designed to help difficult thoughts move instead of harden.
Learning experience
Interactive browser missions that teach Ada and SPARK through small systems bugs, compiler-style feedback, and proof.
Learning experience
An interactive IBM mainframe HLASM simulator with guided missions, a two-pass assembler, virtual CPU, EBCDIC, packed decimal math, and time-travel debugging.
Human ↔ agent network
Where humans and agents find each other, signal what they need, and show what they can do.
Community project
A larger bet on making Ada delightful again with an in-browser proving ground, SPARK, and an MCP kit for AI coding assistants.
Multi-model harness
Where an LLM panel argues before going on stage — blind debate, scope checks, and submission sentries, with you keeping the ledger.
Agent debugging
Find the memory that broke your agent — inspect what it learned and which recollection sent it off the rails.
Reading experience
One topic, five editorial perspectives — no feed, no algorithmic addiction loop.
Built end to end by Rune, my engineering agent, as its first project.
Generative music
A concert-hall playground for generating original music inspired by classical composers while learning how their styles work.
Built as an experiment with Grok 4.5.
Agent-to-agent connector
An unofficial concept, not affiliated with or endorsed by Sentry or Meta: a small connector that lets Meta's Muse assistant search and triage Sentry issues with your own token, testing agent-to-agent integration out of the box.
That is usually where the useful conversations begin.