AI that helps, and knows when to ask.
Most teams do not need a chatbot. They need a system that reads a request, finds the right information, prepares an action and knows when to ask a person. AI is the newest part of my work; the habits behind it are not. I design these systems the way I built trading software: the model makes a bounded choice, code checks it, and nothing irreversible happens without approval.
First step A small pilot on your real cases, measured, with an honest verdict on whether it is good enough.
What this looks like
- Plain-English questions over your database or documents, with answers you can trace back to the source.
- An assistant that drafts replies or actions from incoming requests and waits for approval.
- An AI feature inside your product, with a test set that catches regressions when the model changes.
- A review of an AI prototype that worked in the demo but not on real data.
A good fit if
- A repetitive process needs reading, judgment or writing, not just rules.
- An AI prototype worked in the demo but not on real data.
- You want AI inside your product and need it to behave predictably.
- Your data cannot leave your infrastructure, so local models are worth considering.
Probably not a fit if
- You want a general chatbot with no connection to your systems.
- The task follows known rules. Then plain code is cheaper and more reliable, and I will say so.
How I approach it
-
The right amount of AI
Plain code where rules are known, a small decision model for narrow judgments, a large language model where reading and writing matter, and a person before anything irreversible.
-
Only valid options
The system offers the model choices it has already checked against your data and permissions, so it cannot invent a customer, a file or an account.
-
Approval before actions
Anything that writes, sends or deletes can be previewed and confirmed. Every step is logged, so you can see why something happened.
-
Measured, not assumed
A test set built from your real cases shows how often it is right, how often it refuses, and what changes when you switch models.
Evidence
Questions
Have you shipped AI for clients before?
Not yet for a paying client; my AI work so far is open source, so you can read it before deciding. That is why the first step is a small, measured pilot on your real cases, with a plain verdict on whether it is good enough.
Which AI models do you use?
Whatever fits: hosted models when quality matters most, local open models when data must stay in-house, and small decision models for narrow, repeated choices. The design avoids depending on one vendor.
What stops it from making things up?
The model is not trusted on its own. It chooses among options your system has validated, code checks every value before use, and unsupported requests are refused rather than guessed.