Most case studies hide the messy part. This one starts there.
PokerInk is where we test Agentic Engineering under real constraints: small team, changing requirements, and no luxury of long feedback cycles.
The objective was not "adopt AI." The objective was tighter execution: shorter path from code to evidence to decision.
Context and baseline
PokerInk is a poker training and hand-journaling product: a Next.js web app, a Hono API on AWS, and native iOS and Android apps, all in one monorepo since April 2026. It started as three separate repos with three separate release processes.
The team needed a reliable way to ship quickly without losing operational control. Traditional handoffs were too slow and too optimistic. Problems surfaced late, and weak assumptions stayed alive too long.
We treated this as an operating-system problem, not a tooling problem.
Under it, we run two solid foundations: one for infrastructure and one for application delivery. The Infrastructure Boilerplate and Next.js Boilerplate are built by experienced engineers, optimized for AI workflows, and shaped by decades of practical lessons.
What runs today
Delivery path: push to main with automated gates. Every push runs 41 Playwright end-to-end tests, type checks, and secret scanning before anything deploys. Web, API, and mobile ship from the same pipeline. In the last 30 days that produced 62 production deploys, about 2 per day, with no manual release process.
Mobile path: JavaScript changes ship over the air as staged canaries. A new update reaches a fraction of users first, and only ramps to everyone after it looks healthy. A bad release cannot take out the whole install base. Native changes go through store builds, and both apps are live in the App Store and Play Store with in-app subscriptions.
Production path: daily monitoring and feedback through PostHog and Sentry, plus a daily agent run that reviews open PRs, checks costs and production health, and collects compliance evidence.
Decision path: challenge major bets before execution so the team can kill weak plans before they become expensive work.
Rafiki ties these loops together operationally by surfacing signal, tracking context, and reducing coordination drag.
What broke and what changed
Not every decision worked. A visible example was Instagram strategy: we ran a content approach that looked plausible and failed in live response. We shut it down instead of defending it.
The pipeline breaks too. About 1 in 5 deploy runs fails a gate. That number is not embarrassing, it is the system working: failures surface in minutes, in CI, instead of in front of users. A recent example: our E2E tests were triggering real transactional emails that bounced at the provider. The provider signal caught it, and the fix shipped the same day.
That operating model sharpened with each failure. Now high-impact choices get challenged earlier, and execution starts only after assumptions are stress-tested.
The key shift was cultural: confidence no longer counts as proof.
How it runs now
The rhythm is daily, not weekly. Features land on web and mobile the same day, the agent reviews what shipped, and production signal decides what happens next. We track what changed, what signal moved, what failed, and what gets cut.
This is still evolving. That is expected. The difference is we now have a system that learns faster than it drifts.
Why this case matters
PokerInk is not a polished benchmark story. It is operational proof that Agentic Engineering can run in a real product environment with real tradeoffs and real mistakes.
If your team needs the same outcome, start with loop design, not tool shopping.
Related reading: How we operate PokerInk week to week