01 / The problem
Challenge
Generating implementation code safely requires more than asking an agent to write code. The workflow needs a plan, domain-specific constraints, an automated review loop, and a human gate before changes are promoted.
02 / System view
Architecture at a glance
A simplified, public-safe model. Select a layer to inspect its responsibility.
Selected layer
User story
A bounded engineering request defines the desired behavior and acceptance context.
02.5 / Workflow visual
From story to promotion
A public-safe illustration of the code-generation and review loop. The real client implementation remains confidential.
Original illustration based on a public-safe description. No client systems, prompts, repositories, or standards are shown.
03 / Engineering decisions
Approach
- Build a code-generation agent that works from an implementation plan produced from a user story.
- Steer generation toward secure, compliance-aware coding practices appropriate for a regulated financial-services environment.
- Run an orchestrated code-review loop: inspect deviations, feed findings back to the code-generation agent, re-review, then hold for human approval before promotion.
04 / Tools & skills
What I used
Tools
Skills
05 / Evidence
Outcomes
- Contributed to a growing AI Factory capability.
- Connected planning, code generation, review, correction, and promotion into one workflow.
- Kept the public architecture abstract to protect client and internal details.
Public case study is intentionally sanitized. The client is described only as a financial-services enterprise.
06 / Reflection
What I learned
Boundaries make agent workflows easier to evaluate.
A review loop turns code generation into an iterative engineering system.
Human approval remains essential at the promotion boundary.