Does our data leave our infrastructure?
Not if you don't want it to. The assistant can run on a local model inside your own environment; nothing is sent to a third party.
The problem
Everyone wants AI on the data. Most of it answers confidently, and wrong.
How it works
Three steps, in order.
- Readiness.I check whether the metric layer can carry an assistant, and collect the questions your team actually asks. If the foundation isn't there, that comes first.
- Tools and evaluation. I write the tools the agent is allowed to use, and turn real questions into evaluation cases with known answers. The assistant has to pass them before anyone sees it.
- Ship and watch. The assistant goes where the team works, Slack or the browser, with every answer logged and its accuracy measured against the cases as the data changes.
What you end up with
- An assistant on your warehouse. It answers from the metric layer, not from the internet.
- A documented tool set. Every way the agent can touch data is written down and small.
- Evaluation cases. Real questions with known answers, replayed whenever anything changes.
- The local-model option. The whole thing running inside your own infrastructure.
Proof
I run my own tool this way.
TrueLift, the forecasting and decision tool I built, has an analyst inside it. The agent reaches the data only through purpose-built tools, runs on Claude or on a local model, and every answer carries its working. The scorecard on that page is the same discipline applied to my forecasts.
Who you'd work with
I'm Martin Guzman. I've spent twelve years building marketing measurement: warehouses, attribution, forecasting, and production machine learning. I currently run analytics for the marketing team at a B2B AI company. You work with me directly, with no account managers and no handoffs.
The offer
TODO offer name, e.g. Analytics assistant pilot
You get
- A readiness verdict on your metric layer, and the fix list if it isn't ready
- The ten questions your team asks most, as evaluation cases with known answers
- A working assistant on one of them, in Slack or the browser, with its working shown
- The running cost, hosted and local, before anything else is built
Book a quick discovery call
Questions
Which model do you use?
Claude where the team allows it, a local model through Ollama where it doesn't. The tools and evaluation cases are the same either way, so switching later is cheap.
What if our data isn't ready?
Then the assistant would answer confidently and wrong, and I won't build it. We start with the data foundation and come back to this layer when the metric layer holds.
Can it be wrong?
Yes, and the point of the build is to make that visible and rare: every answer carries its working, and the evaluation cases are replayed whenever anything changes.
What does it cost to run?
A local model runs on hardware you own; a hosted model costs cents per question. The plan states the running cost before anything is built.
What comes first
Start with a call.
An hour with your data and your team's questions. You'll leave knowing whether an assistant can be trusted on them yet.