Consulting
Get an experienced review of the AI decision in front of you.
Work with Sunil Mathew on one use case, architecture or next-step question. Agree the boundary, examine the evidence and leave with a clearer decision to make.
Bring the question that is holding up the work
You may need to decide whether a workflow needs an agent. Your team may have an architecture ready for challenge. Or several possible use cases need a more disciplined way to choose the next investigation.
Bounded advisory focuses on that decision. You remain responsible for what your organisation chooses; Sunil brings an outside view of the technical reasoning and operating consequences.
Three kinds of work
- Frame a use case
Establish the task, constraints, expected value and evidence needed before choosing an approach.
- Review an architecture
Examine the design's assumptions, tool permissions, evaluation, failure handling and cost. Identify where evidence is missing or a decision needs reconsideration.
- Define the next investigation
Turn uncertainty into a useful piece of work: what to test, what to inspect and who needs to decide afterwards.
What a review actually asks
An illustrative case: a returns assistant reads a policy and recommends a refund, and the team wants to connect a tool that issues the refund. On the diagram it is one more box; in production the system can move money. No customer incident is claimed, but these are the questions a real review works through, in this order.
- Does this need an agent at all?
If the steps can be written down before the request arrives, a workflow is cheaper to run, easier to test and easier to defend.
- What evidence has to hold before the action?
Named, listed and checked by the tool itself, not asserted in a prompt and hoped for.
- What is each tool allowed to do?
One job per tool, the narrowest scope that does that job, and limits the tool enforces rather than limits it is told about.
- What happens when the evidence is uncertain?
Three outcomes rather than two: act, refuse, or hand over with the specific gap named. A confidence threshold is not a plan.
- Can the same action happen twice?
A timeout with an unknown outcome is the ordinary case, not the edge case. The system needs a way to examine the state before repeating anything.
- What can you see afterwards?
Which evidence was used, what each tool did, and whether the run stayed inside its boundaries. Without that, an evaluation cannot point at a design change.
The six questions are written out in full, with the reasoning, in the published guide.
What the scope conversation settles
The question to answer, the material to review, the people involved and the output you need. Fee, timing and scope are agreed before the work starts.
This service covers a bounded review or recommendation. Implementation, ongoing AI leadership and an organisation-wide diagnostic require their own scope.
What advisory is not
- Not delivery
No implementation is committed, guaranteed or included. If a review leads to work you want built, that is a separate conversation with a separate scope, and it does not begin by default.
- Not a package
There are no tiers, day rate or minimum term. The scope and the fee are agreed in the scoping conversation.
Why Sunil
Sunil Mathew brings 26 years of engineering experience and past roles at Google, Amazon and Walmart Labs. His current agentic-AI product work and teaching keep the conversation close to the choices engineers and leaders have to explain.
When a different service fits better
- Need a senior owner for the continuing AI agenda? Fractional CAIO.
- Need a diagnostic across the organisation? AI Readiness Assessment.
- Need the team to develop its own capability? Enterprise learning.
Discuss your question
Send a short description of the decision, what you have already tried and who needs to act on the answer. Keep confidential records out of the first message.
apply@thelivingcraft.ai