← All posts Case study

AI in the Dev Loop: Small Team, Wide Surface, Weekly Releases

· 5 min read

A long-lived platform with thousands of users. A team of six. Multiple features shipped per week, across a codebase that's been live for years. The bottleneck wasn't ideas — it was the cognitive cost of moving fast on old code without breaking what works.

Here's how that ran in practice.

The product

A long-running network platform for a large open-source organization — developer teams, founder-led startups, investors, the broader community, all on one surface. Members find each other, share context, communicate, collaborate. Multi-year codebase. Mature core flows. Ambitious roadmap.

The team and the scale

Every account on the platform represents a company, not an individual. This is a B2B network — every member is an organization with real economic weight behind it. That changes the math on user value, and the cost of any regression that affects logins, search, or trust.

The member base roughly doubled over the past year — from ~1,300 organizations to ~3,200 — while the team stayed the same size and shipped weekly throughout. That ratio (six people, B2B platform with 2.5× growth, weekly features on years-old code) only works if the team can move through unfamiliar parts of the system without paying a discovery tax every time.

The lever: AI in the dev loop, not in the product

The accelerant wasn't "we shipped an AI feature." It was that the team embedded AI across the development lifecycle:

Onboarding is the part that matters most for a small team on a wide platform. When any engineer can ask the system "what does this module do, and why" and get a grounded answer, the ramp time for touching a feature outside your usual area drops from days to hours.

A high-ambiguity flow, built in roughly one month

The team built an end-to-end demo-day event flow from scratch in approximately one month. Requirements were unclear at the start and shifted during delivery — classic high-ambiguity work, the kind that usually blows past timelines as the spec stabilizes.

It shipped on schedule. First day of the event drew approximately 3,000 concurrent users at peak — an order of magnitude above the platform's normal traffic — and the system held.

What "AI in the dev loop" actually looks like

A few examples from the team's working practice:

The last point is worth dwelling on. The team has an AI-assisted component library driven primarily by the designer with light engineering support. That's a signal: design–dev collaboration is changing shape, not just "developers using AI tools."

Sustained velocity, not a burst

The hard thing isn't going fast for a week. It's going fast every week, on a long-lived codebase, with a small team, while refactoring as you go. That requires the AI in the loop to be reliable enough that the team trusts it for routine work — and saves human cognition for the things AI still can't carry: architecture under uncertainty, judgment calls on tradeoffs, social complexity, and the kind of bug that lives across three modules.

What this looks like for your team

If you're running:

— the lever isn't headcount. It's embedding AI into the development workflow far past "we use Copilot." It's a meaningful operating change, not a tool adoption.

We did this as an embedded team. Six people. Weekly features. Years-old code. No collapse in quality.