Projects · Mariya Valui
AI systems for business operations
Clear goals, parallel work and results that can be checked.
I start with the business decision, then work out how AI agents can help reach it. This project brings that approach into an operating system for business work: clear goals, parallel execution and results that can be checked.
My contribution
Business problem, product requirements and acceptance criteria, including decision authority and evidence of completion.
Status
In development.
The problem I am working on
An agent can produce a convincing answer while leaving the actual task unfinished. Work can also lose its context when it moves between agents or sessions. I want the system to keep the goal, relevant sources and completion criteria connected throughout the work.
My role
My contribution is the product framing: defining what the system should achieve, what information it needs and what would show that a task is complete. I also define where the system can act autonomously and where an action needs approval. This gives the agents room to choose how to solve the task within a clear scope.
How I think about the workflow
Parallel work is useful when each agent knows the outcome it is contributing to. Research, competing solution attempts and verification can happen together. Their outputs need a common reference point so that a review evaluates the same task and version as the work itself.
From a business goal to a checkable result
- Define the goalState the decision or deliverable and the criteria for completion.
- Work in parallelInvestigate sources, develop alternatives and challenge assumptions.
- Bring the evidence togetherKeep the source and relevant version attached to each result.
- Review the outcomeCheck the completion criteria; bring any action requiring approval with a concrete result to review.
What an alert needs to tell us
An alert is useful when it carries enough context to decide what to do next. The product and source tell us whether it belongs to the task. Its version tells us whether an existing review still applies. In the example below, the alert is relevant—but the review needs refreshing.
Illustration of the mechanism
Alert: review needed
- Product
- Operations
- Source
- Alert registry
- Alert version
- Version B
- Reviewed version
- Version A
Check the match
Does it belong to this task? Yes: the product matches the scope of the current task.
Has this version been reviewed? Not yet: the alert is for version B, while the saved review covers A. The current version needs checking.
Illustrative example: names and versions are demonstration values.
What this work is teaching me
The useful product question is often more precise than “Can AI do this?” It is: what does completion mean for this task, and what evidence would let another agent or a business owner check it? That is the question I keep returning to as I develop the system.
Read my analysis of verifiable AI work ↗
Design example informed by the internal Core development report, 4 October 2026: alert-source scope and release identity; local release candidate.