LLMs for Engineering Alignment and Productivity

Jon Evans and I worked together for years at HappyFunCorp, helping grow it through its eventual sale to Tiny Capital. After Jon moved on – to Metaculus, where he was following LLMs closely – we kept comparing notes, and the more I built with these models the clearer it became: this was going to be a transformational force in software development, one that would put a premium on judgment and strategy. My time at HFC had run its course after the sale, and I wanted to spend it hands-on, building with and for these tools. We landed on a shared thesis: efficiency and transparency for product engineering teams, powered by LLMs. After working with dozens of teams with our time together, we felt there was a significant role to play for LLMs to help improve alignment.

Dispatch AI pulled from GitHub, Jira, Linear, Confluence, Slack, Google Chat, Figma, and more, to produce a weekly alignment report for each project: progress, risks, and detailed dives that went all the way down to the code. We started by reviewing open-source projects by hand to learn what a good report even looked like, and ran a dogfood report on ourselves to tune the approach. We were deliberately not in the action-automation business – what we saw from teams further ahead was too error-prone and fragile to trust at that time.

The bigger ambition was connecting the golden context of a business – its actual goals – to what was happening downstream in the work, and incorporating signals from analytics and even the ERP. That's something the mainstream is still without (Cowork is probably the closest analog today), and it turns out to be as much an org-design and operational-discipline problem as a technical one. Maybe someday the Dispatch dream becomes reality; no one has really gotten there yet.

We grew to six paying customers with interest in more build-outs, but couldn't get traction with the investors we'd need to keep going – and having gone full-time on it, my runway was real. Jon had architected the system, so we open-sourced the backend – YAMLlms, a YAML-driven framework for composing LLM tasks – to show the work.

What I carried out of it was helping in how I approach building today: coding daily as the tooling matured, learning how LLMs actually behave across different tasks – how to simplify a problem, build evals, design workflows – and an equally deep sense of what they're wrong for, despite pressure to apply them anyway. We also lived the adoption side firsthand: the cultural resistance, the data-security conversations. Those hard-won hours have been instructive in the work I continue today.