- My contribution
- Application architecture, implementation, and governed agent workflows.
- Implementation
- TypeScript, Next.js, Fastify, PostgreSQL, workers, and scoped tools.
- Evidence
- Source checks, authenticated interface verification, and a fictional pilot. General production adoption is not claimed.
- 01Context
Shared business records
- 02Proposal
A scoped agent action
- 03Review
Independent quality checks
- 04Approval
A person and an exact version
The conversation needed a system behind it.
A chat interface can make a complex operation easier to use, but the conversation should not become the only place where important business state lives.
I worked on an operations platform where shared records, tasks, documents, approvals, and audit history remain in the application. AI is one way to work with that information, with a defined scope and a clear person responsible for the next decision.
Keep authority in the application.
- Shared records as the foundation
- Contacts, companies, opportunities, tasks, and documents belong to a persistent data model. Agent runtime memory is replaceable.
- Scoped access
- Actor and company authorization are enforced by the application, with company isolation in the database and scoped tool interfaces.
- Version specific approval
- A review applies to a particular artifact version. Source references and checksums help keep the decision attached to what was actually reviewed.
- Independent quality review
- Preparation and quality review are separate roles. A proposed artifact must pass the relevant checks before it can proceed.
A pilot that exercised the boundary.
The recorded implementation checks included automated TypeScript tests, package type checks, linting and builds, and authenticated interface verification.
A fictional case was also exercised through the workflow. It stopped at the independent quality review gate, demonstrating that a prepared artifact did not automatically become an approved result.
A foundation for accountable AI work.
The delivered engineering scope was an implemented fictional pilot with shared business state, scoped actions, evidence, and approval history. It showed how AI assistance can sit inside an application that retains control of permissions and decisions.
This case does not claim general production adoption, final client acceptance, or authorization to process real customer or legal records. Moving beyond the pilot requires a separately agreed rollout, operating checks, and approval of the actual use.
Written by Younes Nadif
