Taking AI output through review and execution

Making generated work usable

Engineering teams need requirements, designs and test cases that reflect their project. They also need to inspect the output, correct it and follow it through execution.

I implemented capabilities across the backend, interface and integrations of an internal AI engineering platform: document and repository context, agent tools, artifact review, diagrams, exports, publishing and QA execution.

Keeping project context with the work

I added retrieval from authorized documents, tickets, repositories and saved project work. This gives agents relevant material for a request and keeps the resulting artifacts connected to subsequent reviews.

The difficult part included access at execution time. I investigated sessions where a tool was configured but could not read the intended repository, tracing the gap between the agent’s available tools and the connector’s actual access.

Making review part of the workflow

I implemented conversational refinement and artifact review so users can clarify requirements and correct interpretations before publishing. This supports a practical cycle: draft, discuss, revise and review.

I added reviewed workspace feedback to the guidance applied to future agent work and implemented reuse of existing parent tickets during Jira publishing to address duplicate epic creation.

These decisions keep user corrections attached to the work and make ticket publishing aware of the existing project structure.

Connecting test generation to execution evidence

I added test generation and reuse, repository-based execution, connector-specific checks, execution-target validation and run tracking. The workflow surfaces failures and supports publishing test results through configured integrations.

I investigated authentication mismatches, target-verification gates and CI failures across the platform and its automation workflow. This required following the request from the agent to the runner and the target system.

Making the platform inspectable

I implemented rules-based model selection and usage/cost visibility, alongside administrative views of agent activity and execution. This gives operators a way to inspect what ran and the resources it used.

My work also covered failure handling and the user experience around long-running work, including surfacing execution status and results.

Implementation stack and deployment

Backend and agents

Python, FastAPI, Google ADK, Gemini

Interface

TypeScript, React, Next.js, CopilotKit / AG-UI

Deployment

Docker, Google Cloud Run, GitHub Actions

Stage

Internal production use, with ongoing operational improvements

Scope and constraints

• Connector operations require configured access, supported systems and valid execution targets.
• Workspace feedback updates agent guidance; the underlying model is not retrained.
• Routing selects among configured models. Output quality still requires evaluation and review.

This account describes my implementation work in general terms. Company source code, internal architecture, customer data and operational metrics remain private.