Cost
100 to 10,000 times cheaper per answer
Analyzing every machine, market, or customer every hour stops being a budget request and becomes a setting you switch on.
Small · Fast · Checked
Expert-level analysis of everything you measure, all the time.
HeartCut teaches a small AI model to answer your recurring questions about your data the way a leading ("frontier") AI model would. The small model runs on an ordinary computer for a fraction of a cent per answer, and every number it writes links back to the record it came from.
What you get
Four benefits from one design
A small model, trained to answer a fixed set of questions, on examples that were checked before it saw them, running where your data already lives. Each benefit stands on its own, so you can adopt HeartCut for just one of them.
Speed
Under a second for a one-question specialist
Answers arrive while they can still change what happens: during the shift, the trade, or the race, not in the review afterward.
On-site
An ordinary CPU where your data already is
No outside service to call or special hardware to maintain. It can run on a plant network with no internet connection, on a bedside monitor, or on a small on-site device.
Checkable
Every number traces back to its source
Each number in an answer points to the record it came from, and software checks the pointer. Someone who was not there can audit the answer afterward.
The name
Keep the heart. Cut the rest.
A distiller runs a whole batch and keeps only the heart cut: the clean middle, after the harsh heads and the heavy tails are drawn off.
HeartCut does the same to a leading AI model's reasoning. The large model teaches, but nothing it writes is trusted automatically. Every answer is checked against the underlying data, and only reasoning backed by the data reaches the small model.
CutHeads
Made-up evidence. A number that cannot be found in
the source record is rejected, however convincing the writing around it sounds.
KeptHeart
Grounded findings. Interpretations, explanations, and
recommendations in which every number points back to the data. The small model learns from
these.
CutTails
Broken or contradictory answers. Answers that ignore the
required format, or that contradict themselves or the record.
The economics
What a year of constant analysis costs
Take 20 questions about 1,500 machines, asked every hour: 720,000 answers a day. Every API figure below already assumes the cheapest practical setup, with questions batched together and the record cached.
- Large-model API
- HeartCut
Figures are modeled from September 2026 list prices. A token is a word or part of a word. The one-time build, assumed at about $500, is not included. Assumptions and caveats
Where it fits
The same questions, asked of everything, all the time
HeartCut is for recurring analysis of organized, mostly numeric data, such as sensor readings, financial figures, and logs. You ask a fixed set of questions, and the answers change as the data does.
- What changed significantly this week?
- Which metrics are behaving abnormally?
- What appears to be driving the change?
- Which markets are diverging from their peers?
- Where are supply and demand becoming unbalanced?
- What deserves attention right now?
How it works
From teacher to student
A large AI model teaches, an automatic check filters what it writes, and a small model learns from what passes.
See the steps Why HeartCutThe finding nobody went looking for
Leading AI analysis is saved for the big incident and the after-action review. The cost is everything nobody looked at.
Read the case How it's differentExpert answers you can check
Each existing kind of tool delivers one half: expert-level findings or checked numbers. HeartCut delivers both.
Compare approachesThe claim
The question changes from "which of our data is worth analyzing?" to "what are we missing by not analyzing all of it?"
A leading AI model teaches the reasoning.
HeartCut checks it and boils it down.
Small models run the recurring analysis at scale.
People act on the findings.