Shows its working.
How an engagement runs, what the language models do and do not do and how your data is handled. Written so that you, or your accountant, can check each claim.
4 steps, every run
Connect
read-onlyTervisi connects to the systems that already hold your sales, orders and costs, with read-only access. Nothing is written back, nothing is changed and nobody has to export a spreadsheet each week. If a system has no connection that can be made safely, you are told before anything is quoted.
Analyse
scriptedThe analysis is done by Python and R scripts that run on a fixed schedule. The language models decide what to investigate and write those scripts, and the scripts compute every figure. Your cost treatment, customer hierarchy, seasonality and definitions are set down once, in code, and used every time. A finding is never a figure someone typed or a model guessed.
Recommend
versionedEach finding comes with what to do about it and what it is likely to be worth, or an explicit no action when the movement is noise. Recommendations are kept with the version of the code and data that produced them for as long as the engagement runs, so you can see later exactly what was said and why.
Account
a receipt for every runEvery run leaves a record: which scripts ran, against which data, at what time and what they produced. If a figure is ever questioned, the record shows how it was made. It is the same record that lets you check the work without taking anyone at their word.
No unscripted numbers
Every figure in a Tervisi report is computed by code, not generated by a language model. Language models orchestrate the pipeline and draft the narrative. The narrative is then checked against the computed values before you see it.
What the language models do
Decide what is worth investigating, write the scripts that investigate it and draft the prose that explains what was found.
What the code does
Computes every figure. While the engagement runs, the scripts are kept with the report and can be run again on the same data, by anyone you choose.
Where the check sits
Between the draft and you. Each figure the prose mentions is compared with the value the code computed, and a report that does not match is not sent.
Figures reproduce. The story is a recorded judgement.
Run the scripts again on the same data and you get the same figures. The story told about them is different: choosing which movement matters most is a judgement, not a calculation.
So the pipeline drafts several candidate stories and chooses one. The ones it rejected, and the reasons, are kept. If you disagree with the emphasis, you can see what else was considered.
What this does not claim
A trace tells you where a number came from. It does not tell you the number is right. Those are different claims, and only the first is made here.
A figure can be computed correctly from source data that is wrong: a cost booked to the wrong product, a return never recorded. That is why every trace ends at the rows themselves, so the data can be checked, not only the arithmetic.
The receipt
After each run you get a short summary by email. It records:
- recommendations what was recommended, and why
- models which language models were used
- calls how many times they were called
- tokens how much text they read and wrote
- compute how many hours of computing the run took
How your data is handled
On one screen, so it can go straight to whoever looks after your data.
- Access
- Read-only connections. Nothing is written back to your systems.
- Models
- Nothing is pasted into a public model. The models run under zero-data-retention terms, and nothing trains on your data.
- Retention
- The model providers keep nothing after a run. Tervisi keeps only what the reports and their traces need, for as long as the engagement runs.
- Deletion
- On request at any time, and in full at the end of the engagement.
Book a call
30 minutes on Zoom. Free. You leave knowing whether your cost data can support margin analysis.