Most AI tools guess.
Quantiri computes.
The difference between Quantiri and a chatbot pointed at a spreadsheet is architectural. Every number in a Quantiri report is computed by a deterministic statistics core before any language model runs. The model writes the story — it never invents the facts.
Two systems, not one.
Most AI analytics asks a single language model to do everything: find the patterns, do the math, and write it up. That is exactly how you get confident, fluent, and wrong. Quantiri separates the two jobs that should never have been combined.
Deterministic core
Statistics are computed by code — measured, repeatable, auditable. The same dataset always yields the same numbers. No model is involved in producing a single figure.
Constrained narrative
Only after the facts exist does a language model write the report — and it is given the computed evidence and nothing else. It interprets; it does not calculate.
Four phases.
Three of them never write a word.
The first three phases are pure computation. Only the last involves language — and by the time it runs, every fact already exists. This ordering is the whole game.
Understand the shape before the contents.
Before any analysis runs, Quantiri reads the structure of your data — what each column is, what role it plays, and what should be ignored. Identifiers and noise are excluded so they can never masquerade as findings.
Patterns are found by statistics, not by a chatbot.
Specialised detectors search the data for genuine structure and only surface a pattern when it clears real statistical thresholds — both that an effect exists and that it is large enough to matter. Detectors run independently, so a weakness in one never corrupts the rest.
Ordered by what would change a decision.
A deterministic selection step keeps only the handful of findings that most deserve a decision-maker's attention, and balances them so a single kind of pattern can't crowd out everything else. The most important finding is always first.
Language is applied last — and only to the evidence.
The writing model receives the computed evidence and is constrained to it. Its job is to explain, in the voice of a senior analyst, what the numbers mean and what to do next. If a figure was not computed upstream, it cannot appear in the report.
Zero hallucination.
Not a claim. An architecture.
“The AI won't make up numbers” is a promise most tools can't keep, because their numbers come from the AI. Quantiri's come from code. The trust isn't a feature we added — it's a consequence of how the system is built.
Every statistic in a report is computed before the language model runs. The model is handed the numbers — it is structurally unable to make them up.
When the report derives a figure — a ratio, a gap, a projection — it shows the arithmetic it used, so any reader can check the math against the evidence.
The writing step is confined to the computed evidence. A claim that isn't backed by an upstream measurement simply cannot enter the report.
Each report states, in plain language, that its figures came from the computed evidence and nothing was imputed beyond the arithmetic shown.
Built like infrastructure,
not a prompt.
The interesting part of Quantiri isn't that it uses an LLM — everything does. It's the system around the LLM that decides whether the output can be trusted.
The analytical core is pure computation. Identical input produces identical findings, every run — a property most LLM products cannot offer.
Detectors and stages are independent. One stage degrading or failing never takes down the report; it simply contributes less.
Nothing surfaces on fluency alone. Effect size and significance are required before a pattern is ever considered a finding.
Tuned for the 50–500k row range that real business data lives in, with a finished report in minutes rather than hours.
A single typed schema flows from the analysis engine through the API to the rendered report, so the contract is checked, not hoped for.
A dedicated analysis service does the statistics; a Next.js application does the experience. Each can evolve without destabilising the other.
What it deliberately
doesn't do.
Quantiri analyses numbers and categories. Free-text corpora and binary/image data are outside its current scope.
Meaningful findings need a real sample — roughly 50 rows or more. Below that, there isn't enough signal to be honest about.
Quantiri surfaces and explains what's in your data. It doesn't decide for you — the judgement, and the context it can't see, stay with you.
The architecture is the point.
The report is the proof.
Read a complete report generated by this exact pipeline — every number computed, nothing invented — then join the early-access list.
This page describes how Quantiri works in principle. The specific detectors, scoring, and thresholds that make the output good are part of what we're building — happy to go deeper in conversation.