5.6× more development output at Keeps

Remove duplicate work

Keeps’ dashboard, admin tools, and background services were parts of the same application, maintained in separate repositories.

A database change could require updates in all three. The same was true for billing and other shared features: related work spread across codebases and multiple pull requests.

We brought the applications into one repository with shared packages. Shared logic had one home, and related changes could be implemented and reviewed together.

Remove a bottleneck from every release

A finished change still had to move through a chain of branches. Each promotion added coordination and opportunities for merge conflicts, pulling attention away from development.

We removed the intermediate branch promotions from the normal release path. Once a change was ready, merging it into the main branch triggered deployment.

That cut repeated handling out of the workflow. Engineers could spend their time implementing and checking changes instead of moving the same work between branches.

Build the guardrails for faster development

To move fast safely, you need guardrails. In a car, you put on a seatbelt. In software, you build tests that help catch mistakes before they reach customers.

We started with a suite of 457 tests in March, then added browser tests, shared coding guidance, and continuous integration. Tests became part of implementation and code review, especially around sensitive areas.

In the PRs tracked here, 81% included test changes in May. In September and October, that reached 89%.

Put AI to work beyond code completion

Keeps already used AI to suggest code as engineers typed. We introduced multiple assistants that could investigate problems, propose an approach, implement changes, write tests, and review the result.

Shared instructions kept the assistants aligned with the codebase and its conventions. Engineers could delegate substantial parts of a task, then inspect and verify the work.

This also created the foundation for automation: well-scoped bugs could move from investigation to a proposed fix without someone prompting each step.

Give engineers ownership of the next step

Faster tools only helped if the team could act. We changed the expectation from waiting for detailed direction to investigating the problem, choosing a reasonable approach, and taking responsibility for the result.

Engineers gained room to make routine implementation and design decisions. Tests and shared conventions supported those decisions, and questions went to the founder when they needed business direction.

That reduced the need for handholding and repeated conversations before work could move forward. The founder could focus on priorities while engineers owned delivery.

Merged changes rose from 161 in 2025 to 904 in 2026 over the same January–October window. This reflects the combined changes to architecture, practices, staffing, and AI usage.

With AI assistance, we also built browser-based automations that let Keeps interact with upstream providers where usable APIs were unavailable. Those automations provided an API-like interface for the product.

Automated investigation meant a fix could be prepared before anyone started chasing the issue. Extending that workflow to more well-scoped problems offers a way to reduce manual follow-up further.

Median ticket closure fell from 20 days in Q1 to 7 days in Q2. It rose to 12.6 days in Q3 during security hardening and a substantial integration build.