In production · SciPlay
I built and run AI workflows at SciPlay to turn design documents into usable tickets, surface release risks, and prepare delivery reporting.
Two feature buildouts
Building ticket hierarchies, translating design documents, and preparing defensible forecasts took recurring effort. Across teams, aging pull requests and release mismatches needed a clearer view.
I built agents and scheduled reporting around those jobs. The workflows prepare the information and proposed changes. I retain approval of Jira changes and ownership of delivery decisions.
The workflows
Read the design documents together, propose a ticket hierarchy, and populate blank descriptions. Conflicting requirements are flagged for a decision.
The 11 conflicts and documentation gaps flagged across both features were potential rework risks, identified before development.
A report reconciles open pull requests, Jira fix versions, and branch targets, then posts to Slack three times a week. Authors and leads are tagged; uncertain author matches are escalated to me.
Reconcile and surface items for review
An open PR against a released version prompts investigation. It does not, by itself, establish whether the changes shipped.
Total hours divided by total capacity can hide bottlenecks. I look at discipline queues, dependency sequencing, and the time QA needs inside the sprint. Spare client capacity does not clear a server queue.
The safeguards
Setup questions establish scope. A separate review and approval step comes before Jira changes.
Gaps and contradictions are flagged, with the decision owner identified. The rule is to surface uncertainty rather than fill it with a plausible specification.
Inventory the tickets first, write into blank descriptions, and compare existing content afterward. On the larger feature, seven existing descriptions were left untouched.
Read the fields back and check the result in code. A successful API response is only one part of verification.
Source reads use MCP; approved Jira changes use the Jira API. Work is keyed on stable IDs. Persistent memory supports retrospective comparisons across sprints. I also wrote the prompt governance standard used by these workflows. Implementation details remain with my employer.
The ticket buildout also surfaced conflicting requirements. The release report brings discrepancies into view so an owner can investigate. Those are the decisions the workflows are designed to support.
Let’s talk about delivery operations, release risk, and team leadership.