Why HeartCut
The costliest finding is the one nobody went looking for.
Leading AI models are excellent at explaining what measurements mean. But every answer runs the whole giant model, so they are saved for special occasions.
The problem
Expert judgment goes to the big incident, not to everything else
The most valuable step in any analysis is going from "chilled water flow fell to 8.2 m³/h" to "the batch failed on cooling, not on the solvent lot; check the chilled-water header before the next batch."
A leading AI model is the most capable tool for that step, so it gets used sparingly: on the incident but not on every machine, on the session review but not on the next lap. The insights that running it on everything, all the time, would have produced never exist.
Why it costs what it costs
It is expensive because it knows everything
The same model that can reason about bearing wear also holds Ming dynasty ceramics, the tax code, and the rules of chess. Its price and response time follow its size, not the difficulty of your question.
A business that asks the same handful of questions about one domain rents all of that capacity and uses a sliver of it. All that general knowledge is worth paying for, but not on every answer forever. HeartCut pays for it once, while the small model is being built.
Four reasons, each enough on its own
What a small, checked model changes
Cost gets the most attention. For many applications, one of the other three decides the matter.
Analyzing everything becomes a setting
100 to 10,000 times cheaper per answer. Analyzing every market, machine, product, and reporting period no longer needs a budget request.
Answers arrive in time to matter
Under a second for a one-question specialist. Findings arrive while they can still change the outcome: during the shift, the trade, or the race, not in the review afterward.
The data never leaves
An ordinary CPU beside the data, with no special hardware. That reaches places a GPU cluster cannot: a plant network with no internet connection, a bedside monitor, or a small on-site device.
Every number can be defended
Each number in an answer can be looked up in the source record. Someone who was not there when the answer was produced can audit it later.
Cost and speed
The arithmetic of constant analysis
One answer is one question about one monitored item, such as a machine or a market. The workload here: 20 questions about 1,500 machines, every hour, or 720,000 answers a day.
| Option | Per answer | Per year | Time per answer |
|---|---|---|---|
| Large-model API, long record (30k tokens), mid tier | $0.0221 | ~$5.8M | 5–43 s |
| Large-model API, short record (3k tokens), mid tier | $0.0158 | ~$4.2M | 5–43 s |
| Large-model API, short record (3k tokens), low tier | ~$0.0053 | ~$1.4M | 5–18 s |
| HeartCut general model, on a CPU at the data | ~$0.00002 | ~$5,300 | 4–6 s |
| HeartCut one-question specialist, on a CPU at the data | ~$0.0000033 | ~$870 | ~0.8 s |
Assumed cost to build one HeartCut model. The table leaves it out.
Of constant use at this volume repays that build, compared with the API.
The multiple of the top API row's cost if each question is sent alone and nothing is cached.
These figures show what becomes affordable. They are worked examples at September 2026 list rates, with developer time left out on both sides. Nobody pays the top row today, so read the gap as analysis that becomes possible, not as savings on a current bill.
Findings, not figures
What answers look like in different fields
Each is an interpretation or a recommendation, never a simple lookup, written for the person who will act on it. The examples are adapted from HeartCut's research into each field, shortened, and shown without their citations.
"Cooling trouble on B24-0871, not the solvent. Chilled water flow dropped to 8.2 m³/h and the jacket ran hot for 38 minutes. None of the other batches on lot 24K118 ran hot. Check the chilled-water header before the next batch on Train 2 starts cooling."
For the shift supervisor, at handover
"The request is slow because of how we call pricing, not because pricing is slow. We make 24 pricing calls, one per result and one after another, and together they hold 744 ms of the 1,284 ms. Making them concurrent gets this request to 598 ms."
For the service owner, in the ticket
"An outside party most likely used this employee's password. The sign-in came from Singapore on a single factor, 2 hours 24 minutes after one from Lisbon, and 14 minutes later the same address granted administrator access to a service account. The policy that would have stopped it is in report-only mode."
For the CISO
"Software is up $71,905 on August, and most of it is how things were booked, not new spend. $58,200 is an annual renewal that moves to prepaid, and a $19,850 August bill arrived a month late. With the renewal moved, the month is up $13,705, and that still carries the late bill."
For the CFO, at month end
"Protection worked: breakers 2312 and 1182 cleared a B-to-ground fault, and the first reclose at Westgate succeeded, so the line is back in. The crew found a cracked insulator; schedule the repair outage with them before it faults again."
For the system operator
"The time goes in the long right and the hairpin. In the hairpin you come off the brake and wait: the throttle comes in 17 m later than the coach's, on 9 of 11 clean laps. Next time out, carry more speed through the long right, where you're at 97 km/h and the coach is at 104."
For a track-day driver
Where it applies
Exact data, a fixed set of questions
A field qualifies when its data has relationships that can be checked and the questions asked of it come from a fixed, repeating set. HeartCut is designed for applications across seven kinds of data.
| Kind of data | Applications |
|---|---|
| Time series and sensor readings | Predictive maintenance, IT monitoring, intensive care monitoring, energy grids, financial risk, process manufacturing, industrial control, motorsport coaching |
| Networks and linked records | Knowledge graphs, supply chains, application performance tracing, logistics routing |
| Networks that change over time | Analysis of relationships between people and organizations, disaster relief coordination |
| Scheduling and allocation | Workforce scheduling and rostering |
| Spatial data | Weather and climate, genomics, robotic toolpaths, field sports, chess coaching |
| Math and formal logic | Mathematics tutoring |
| Documents with addressable passages | Disclosure and redaction review |
Answers as the product
Run your questions against each new batch of data, covering everything whether or not an alarm flagged it. Use the answers directly in reports, alerts, per-machine summaries, and dashboard text.
Answers as inputs
Because answers are structured and cited, other software can use them: combining them across machines, comparing them over time, or passing them to another rule or model.
Answers for an assistant
A chatbot can ask the model directly, start from answers computed ahead of time, or look them up in a search index when there are too many to fit in one prompt.
Next
Who else does this, and when is HeartCut the wrong choice?
A comparison with the other kinds of tools, including the cases where an API call is still the right answer.