Engineering

Built like a chip, not a chatbot.

A different kind of tool

Not a dashboard. Not a chatbot. Something that looks for you, and can prove it.

The questionTraditional BIAI analyticsOthor AI
Who decides what to look atYou, by building a chartYou, by asking a questionNobody has to: every metric and breakdown is tested every night
Who produces the numbersYour queriesOften a model, writing SQL or the answer itselfCode only. A model never computes a figure
Ask twice, getThe same chartPossibly a different answerThe same numbers, byte for byte, from a replay
How sure is it?Not statedRarely statedGraded on every finding, after false-discovery control
Where a number came fromIn the dashboard’s definition, if you lookUsually not shownDefinition, SQL, test, parameters, spec version and run, one click away
When the AI provider is downNot affectedStopsThe night runs; only the wording falls back to templates

The night

Every night, the same pipeline.No model inside it.

  1. 01ReadRead-only logins · only what changed
  2. 02MaterialiseColumnar files on the data volume
  3. 03ComputeEvery confirmed metric, by SQL
  4. 04Test11 detectors · forecasts · checks for chance
  5. 05LedgerAppend-only · replayable
  6. 06BriefRanked, graded, by 06:00
AI model · outside the pipelineOptional. Drafts the brief’s sentences from computed results; the numbers guard checks every one. Off, and the brief uses templates.

Built lean

Two things stay small: your AI bill and your time.

A model drafts once, while you set up. Code does the rest, every night, with no model at all. And Othor AI asks as little of you as it can: a few steps to start, then a short read and a few taps each morning.

AI bill

Indigo Thread Co., our sample company

Read the documents you uploaded
3 calls
Draft how your business works
1 call
Describe the tables it found
4 calls
Draft your metrics
1 call
AI in the nightly run
None

To set up, once

$0.61

Each morning, per brief

1¢Optional: smoother sentences, numbers unchanged

9 calls, measured 10 Nov 2026 from the calls Othor AI logs, at the model’s list price of $4 and $20 per million tokens in and out.

Human effort

Indigo Thread Co., our sample company

Steps to set up
5
Metrics drafted for you to confirm
29
Morning Brief
150 words
Questions on Home (each a tap)
5
Proposals to accept or decline
1

To set up, once

5 steps

Each morning

1 minTo read the brief, then 6 taps on Home

The morning after an ordinary night of the sample: its brief read at 200 words a minute, and what was open on Home. We’ll measure real owners’ time with our first companies.

Speed

Built to stay fast as your data grows.

Every page

under 120 ms

Ten times the data, and every page still opens in under 120 ms. A whole night of reading, computing and testing takes under a minute, and replays to the same result, byte for byte.

How we measured

205 thousand rows1.54 million rows

Every page, at both sizes

Median server time for each page at 205 thousand and 1.54 million rows; every page under 120 ms, against a budget of 300 ms.0 ms100 ms200 ms300 ms300 ms budgetHome14 msMorning Brief10 msSweep105 msA finding, with its proof106 msOntology92 msA metric, with its history115 msMedian server time for each page at 205 thousand and 1.54 million rows; every page under 120 ms, against a budget of 300 ms.0 ms100 ms200 ms300 ms300 ms budgetHome14 msMorning Brief10 msSweep105 msA finding, with its proof106 msOntology92 msA metric, with its history115 ms

Median server time of 15 loads of each page, in milliseconds, against our budget of 300 ms. Lower is better.

An ordinary night, step by step

Seconds for each step of an ordinary night at 205 thousand and 1.54 million rows; the whole night took 17.8 s at 1.54 million rows.0 s2 s4 s6 sRead what changed0.4 sStore it, by day7.1 sCheck the data0.6 sCompute every metric3.6 sForecast0.7 sRun every test4.1 sWrite the briefs0.5 sSeal the run0.9 sSeconds for each step of an ordinary night at 205 thousand and 1.54 million rows; the whole night took 17.8 s at 1.54 million rows.0 s2 s4 s6 sRead what changed0.4 sStore it, by day7.1 sCheck the data0.6 sCompute every metric3.6 sForecast0.7 sRun every test4.1 sWrite the briefs0.5 sSeal the run0.9 s

Seconds per step of an ordinary night, after a first night that read the full history (62.5 s at 1.54 million rows). Measured 10 Nov 2026 on one machine, a Mac with Apple silicon, by scripts/measure_site.py, which loads the sample into PostgreSQL and MySQL and runs the real nightly pipeline. Model calls during the nights: 0.

Eight rules

Rules the code enforces, not promises in a policy.

Each is built into the product, not left to a setting or a habit.

  1. 01

    AI drafts, code computes

    A model drafts setup suggestions and the brief’s sentences. Numbers are put into text from computed results; model text carrying a number that isn’t in the computed inputs is discarded.

  2. 02

    Same inputs, same numbers

    Pinned libraries, seeded routines, percentiles written out by hand, and an append-only run ledger. Any run can be replayed and must reproduce byte for byte.

  3. 03

    The night needs no model

    The engine and scheduler can’t import the AI layer at all, so the run, the tests and the forecasts can’t come to depend on it.

  4. 04

    Read-only, in place

    A login that can change data is refused at connect. Only what changed is read; results are stored, tables are never copied. Personal columns are excluded at the scan and refused at the model’s door.

  5. 05

    Every figure has provenance

    Every number, finding, forecast and sentence carries its definition, SQL, test, parameters, spec version and run.

  6. 06

    Nothing goes live until you apply it

    Drafts wait. Every owner action and system event is written to an append-only log, enforced by the database, not the app.

  7. 07

    One image, anywhere

    One container, configured by environment and a data directory. The metadata store runs on PostgreSQL in the cloud and SQLite on premise, with nothing that needs a service we host.

  8. 08

    Small bills

    Model tokens are spent only on setup drafting and brief wording. Nightly reads are incremental and results are stored, not tables copied.

The statistics

Methods you can’t prompt your way to.

Signal analysis and statistics, not a model’s guess. Every test, across every metric and breakdown, is corrected together for false discoveries, so a finding is a finding and not chance.

11 detectors, every night

  1. Net hiding gross

    01

    Breaks a total into its parts and tests whether offsetting moves are hiding a real change

  2. Changepoint

    02

    Finds the point where a series settled at a new level, and confirms the shift is real

  3. Forecast breach

    03

    Holds each actual against a forecast band fixed before it arrived, and each goal against where the year is heading

  4. Rising variance

    04

    Watches the spread of a process, not just its average, and tests when it widens

  5. Broken relationship

    05

    Tracks how related numbers move together, and tests when that link breaks

  6. Lead and lag

    06

    Searches for numbers that move ahead of others, and tests that the lead isn’t chance

  7. Concentration

    07

    Measures how much rests on a few names, and whether that is growing

  8. Missing records

    08

    Learns each customer’s or supplier’s own rhythm, and notices when one goes quiet

  9. Process artefact

    09

    Looks for values bunching just under limits and round numbers, a sign of how a process is gamed or rounded

  10. Cross-source contradiction

    10

    Reconciles the same number from two systems and flags the gaps that matter

  11. Bimodal average

    11

    Tests whether an average is hiding two different groups, and describes them

See it on your own numbers. The first brief arrives the morning after.

Othor AI is invitation-only while we work closely with our first companies. Tell us about yours.

Request accessSign in