Aller au contenu
ALTIMETRIADEV & TECH

Data & Artificial Intelligence

The dashboard shows a figure. An executive asks where it comes from; the answer takes two days. A metric you cannot explain does not support a decision — it starts an argument.

Every figure on screen must be explainable — the one rule we never negotiate.

Decision instrument
The catalogue of choices

Four live models: data you can take in hand

They run in their frame and answer to mouse and finger: a stream you cross, a network you light up, a territory you raise, a text you move back through time. A shape and a movement — never real data in the scene.

4 pieces from our design stock
Three.js modelsLive model — drive it here

The engine that answers

Seven families of public sources arrive in disorder; the scene cleans and dates them, weaves the links between them, then gathers them into a twenty-six-faceted solid — the twenty-six chapters of a file. Three facets stay dark: what we do not know is written N.F, never zero.

Who it is forData & AI — any management that wants to know what is actually done with a piece of data before trusting it.

Three.js · WebGL · 2 100 points · 620 liens · 26 facettes · lueur · récit en 5 chapitres · repli sans WebGL

designed on 5 September 2026

Three.js modelsLive model — drive it here

The document that cites its source

Three hundred and sixty pages as a wall; a bar reads them one by one; the useful passages come loose and rise into a summary sheet; every line of the sheet keeps a thread to the page it came from. One box stays empty: the document did not say it, and the sheet does not make it up.

Who it is forDocument intelligence — a firm, a finance or legal department that receives thick files and wants answers that can be checked in a second.

Three.js · WebGL · 360 pages instanciées · 14 extraits · fils de citation · grain · récit en 5 chapitres · repli sans WebGL

designed on 5 September 2026

Three.js modelsLive model — drive it here

The law at its date

Fourteen versions of the same text rise one by one, like strata. A cutting plane stops at the date a contract was signed: the version in force that day lights up, the articles repealed since fade, those added later stand apart. A court decision links to the stratum it actually judged.

Who it is forLawyers and legal teams — reading a contract with the text of its time, not today's.

Three.js · WebGL · 14 strates · 90 fibres · 380 grains · plan de coupe · lueur · récit en 5 chapitres · repli sans WebGL

designed on 5 September 2026

Three.js modelsLive model — drive it here

The database that fits on a desk

Five thousand four hundred rows in a heap sort themselves into columns, one per type; what looks alike compresses; the whole settles onto the outline of a laptop. Then a copy leaves for a server rack, and a pulse keeps both current: two copies, one truth.

Who it is forInformation systems & cloud — a company that believes it takes a data centre to query billions of rows.

Three.js · WebGL · 5 400 cubes instanciés · 4 types · compression · synchronisation · récit en 5 chapitres · repli sans WebGL

designed on 5 September 2026

Each piece opens full size, exactly as it runs; the live models are driven by hand.

The catalogue of choices

What data gives to see

Six readings of one holding: parcels, territory, deadlines, the stream of notices scored as it arrives, a trade's climate month by month, and a market's structure — who holds the ground, and how locked it is. Working mock-ups, kept in reserve: none decides for you, each gives something to read.

6 pieces from our design stock
DashboardsWorking mock-up, in reserve

The opportunity radar

The stream of detected notices, scored 0 to 3 continuously, the score distribution and the monthly trend: sorting happens before reading.

Who it is forA bid team that receives more notices than it can read.

designed on 6 July 2026

DashboardsWorking mock-up, in reserve

The market barometer

Published contracts, estimated volume, renewal share, response rate: a trade's climate read month by month, with its split by purchase family.

Who it is forA management team steering its order book on the trend, not on anecdote.

designed on 6 July 2026

DashboardsWorking mock-up, in reserve

Competition and concentration

A market's concentration index, the holders' shares, the procedure alerts: who holds the ground, and how locked it is.

Who it is forA company that picks its battles by the density of its competitors.

designed on 6 July 2026

DashboardsWorking mock-up, in reserve

Territorial heat

France by département, coloured by activity, the five most active alongside: the territory reads at a glance, before the table.

Who it is forA branch network deciding where to push.

designed on 6 July 2026

DashboardsWorking mock-up, in reserve

Signals and deadlines

Framework agreements coming to term, the window to position, the reminder before republication: ready before the notice, not after.

Who it is forA salesperson who wants to approach the buyer in the right quarter.

designed on 6 July 2026

DashboardsWorking mock-up, in reserve

The land-registry map

Land-registry parcels aggregated into cells: France draws itself, computed in under a minute.

Who it is forA developer, a planner, a local authority that thinks at parcel level.

designed on August 2026

Each piece opens full size, exactly as it runs; the live models are driven by hand.

Intelligence that answers for its answers

Architecture first, then where the compute happens. Each screen reads on two levels: the decision it serves, and the method behind it.

DEV & TECHNOLOGY · AI

AI that is useful because it is bounded, evaluated and observable.

The model proposes. Policies, evidence and people remain in control.

Value comes from a controlled task, not a chatbot placed over data.
ALLOW · REQUIRE APPROVAL · DENY · ABSTAIN
Method

Hybrid RAG, passage-level citations, bounded agent orchestration and continuous evaluation.

  • Dense + lexical retrieval · reranking
  • Authorised context · tenant isolation
  • Allowlisted tools · least privilege
  • Structured outputs · schema validation
  • Evaluation sets · abstention rate · drift
  • Agent / tool / model traces · latency · tokens
GenAI observability is a moving field: our telemetry schema is versioned, so it keeps up.
DEV & TECHNOLOGY · SOVEREIGNTY

Choose where the model runs, what it can see and what may leave.

European cloud, on-premise or isolated environment: one usage contract, controls adapted to each boundary.

Sovereignty is a verifiable architecture, not a location on a map.
HOSTING · JURISDICTION · NETWORK EGRESS · EXIT PLAN
Method

Compatible serving API, model registry, encryption, egress control and deployment profiles.

  • Self-hosted inference · compatible API
  • Model, weights, quantisation and licence
  • Continuous batching · KV cache · capacity
  • Encryption in transit / at rest · KMS/HSM
  • Egress filtering · mirrored package repository
  • Air-gapped status only when the environment is genuinely isolated
We reserve air-gapped for infrastructure with no network egress, demonstrated. When we write it, it is verifiable.

Data put to work, told without the names

An analytical warehouse that updates overnight, a control that keeps yesterday's measure to see the gap, a corpus indexed at its date: what we did, what it produced, and what was hard. Names removed, volumes kept.

TRACK RECORD · REAL ASSIGNMENTS, NAMES WITHHELD

  1. 01

    Our own house first, then the same pattern taken up for organisations whose data lives in tools that do not talk to each other.

    When · Built in 2026, run every evening since.

    What we did

    We built an analytical warehouse and the scheduler that fills it. It starts on its own every evening, step by step. Machine off, it starts at next boot; no network, it waits; a failed step does not stop the others and runs again the next day. Every step writes its own dated verdict, and the evening report separates three states: failure, deferral, and success after a busy machine.

    What it produced

    A daily scheduler whose number of steps is read from the code and never from a document, a dated log per step, and a daily status page regenerated at a fixed path — having to hunt for the right file would already be human intervention.

    What was hard

    Billions of rows held on a desktop machine with sixteen gigabytes of memory. That is not a boast: it is what dictates the architecture — source-by-source processing, columnar formats, and a single write lock, because one reader left open is enough to block an entire night.

  2. 02

    A team running its own platform that learns of incidents from its users rather than from its own dashboards.

    When · Put in place in August 2026, run every evening since.

    What we did

    We replaced the measurement that overwrites itself with one that is kept: yesterday's value is dated and retained, the gap is computed and written into the verdict itself, and a drop raises an alert that asks the only question that settles it — is this data, or a rebuildable artefact?

    What it produced

    A dated history per source, a gap computed at every run, and a daily check that reopens the database to read the date of the data actually there — not the job log, not the file timestamp: the data.

    What was hard

    Before this check, one database lost nearly half its volume in a single night. The jobs ran, every check went green, and it was a human who spotted it, by hand, comparing the previous day's map with the evening's figure. A single value cannot be wrong; two values can.

  3. 03

    Teams that must cite the rule applicable on the day of the events, and to whom a search engine always returns the latest version.

    When · 2026, in production.

    What we did

    We indexed a versioned body of rules so that a search returns the version in force on a chosen date, not the most recent one. The vintage is a dimension of the index, not a filter bolted on afterwards.

    What it produced

    An index searchable by text and by date, kept separate from the warehouse it serves, and rebuildable without touching the one in service.

    What was hard

    The first rebuild deleted the old index before building the new one — and it could not finish: there was not enough disk space. Its own log gives the time of deletion, an hour and twenty minutes before it was stopped. Destroying before knowing whether you can rebuild is not a rebuild: it is a deletion followed by hope. Its replacement measures free space first and builds alongside; it can no longer touch what exists.

No client name is published, with or without their consent. We publish no win rate: it is not measured, and an unmeasured figure is not a reference.

What we do

  • Data engineering: ingestion, cleaning, modelling, historisation.
  • Dashboards and analytics, built on a shared definition of each indicator.
  • AI integration on your own documents and your own data.
  • Automation of repetitive, high-volume tasks.
  • Audit of an existing data estate: what the data really says, and what it does not.
  • Checking the grain before any join: an aggregated table cannot answer a question asked at a finer grain than its own — and it does not say so, it returns a plausible figure.
  • Backtesting a score before it reaches a screen: number of observations, years covered, completeness rate. A score that has not been backtested is an opinion that computes.
  • Dated historisation of every reading and computation of the gap: a single value cannot be wrong, two can — and a drop raises an alarm instead of going unnoticed.

What we deliver

Source mapping & indicator dictionary — definition, grain, issuer, frequency, last collection date
Ingestion & transformation pipelines
Dashboards in production
Document assistant connected to your data, with every answer tied to its source document
Freshness check that reopens the database and reads the date of the data actually present, source by source
Automations & connectors
Quality note: coverage, freshness, limits, and what was discarded with the reason why

The method

01

Qualify

Where the data comes from, how often, at what grain, and where the holes are — a number without its scope is not a figure, it is an argument.

02

Model

One definition per indicator: written, dated, open to challenge. Before any join, we write in one sentence the business action it serves.

03

Present

A screen that answers a question, not a wall of charts — and a missing value is shown as missing, never as a zero.

04

Prove

Freshness measured, gaps flagged, false positives counted as defects of the check itself: a guardrail that cries wolf ends up unread.

WHAT YOU WILL HAVE IN HAND · The Proven Signal

Open one factor: it leads back to source, date, scope and what is still missing.

A signal only passes once it has cleared its filters.

We build signals that lead to action without asking decision-makers to trust an opaque score. The screen separates fact, derived calculation, inference, coverage and information not furnished.

The executive sees what changed, why it matters, which decision is proposed and how far the data supports it. A late source does not disappear: it changes the signal state and names the affected analysis.

Open the factors

Facts, calculations, coverage, uncertainty and action

Data owner, data steward, data product, domain, data contract and versioned schema.

Metric semantics: grain, population, filter, unit, period, numerator, denominator and how missing values are handled.

Event time, processing time, watermark, lag, backfill and reconciliation.

Stable event identifier, idempotent consumer, deduplication, quarantine, DLQ and controlled replay.

Backward/forward compatibility, schema registry and consumer contract.

Measured quality: completeness, validity, uniqueness, consistency, freshness and coverage, with a visible denominator.

Lineage dataset → job → run → dataset, provenance, version and OpenLineage run state.

Semantic layer and metric contract, with no divergent logic between dashboard, export and assistant.

Hybrid dense + lexical RAG, reranking, passage-level citation, an abstention threshold and an insufficient-evidence response.

Versioned evaluation set: retrieval, groundedness, citation accuracy, completeness, toxicity and abstention rate, by use case.

Model, prompt, corpus and policy version, drift, promotion, rollback and change log.

Human-in-the-loop and tool authorisation: agent identity, least privilege, input/output, approval, traceability and emergency stop.

Where this comes from
  • OpenLineage (openlineage.io)Dataset, job and run lineageaccessed 2026-07-30established

Artificial intelligence proposes and cites; the rules, the evidence and the authorised person decide.

THE QUESTIONS THAT SET THE PRICE

What we will ask you before we start.

Four questions we ask on every file in this domain. Each comes from an incident someone paid for — us, or a client before us.

Is a table a source? How many sources, and of what scope?

A number without its scope is not a figure, it is an argument. We document every source by its issuer, its frequency, its grain — the row, the lot, the company — and its last collection date, and we count what we discarded.

If nobody asks it — Four teams all right about four different volumes.

At which grain is the question asked — and can the table answer it?

An aggregated table cannot answer a question asked at a finer grain than its own, and it does not say so: it returns a plausible figure. We check the grain before the join, and we write in one sentence the business action the join serves.

If nobody asks it — A national group listed as a competitor on a two-hundred-thousand-euro painting package.

Is the model backtested, and on how many observations?

A score that has not been backtested is an opinion that computes. Every prediction carries its number of observations, the years it covers and its completeness rate — we compute only on what is filled in, and we say so.

If nobody asks it — A convincing prediction over years where the data did not exist.

What happens if a source drops out?

The check reopens the database and reads the date of the data actually present, source by source; a drop triggers an alert, and the previous reading is kept to compute the gap. Collecting is not connecting: a source fresh in its raw form and absent from the screen is a source thrown away.

If nobody asks it — A robot announcing "all done" while three pipelines are broken.

CAPABILITY SCENARIO — NOT A CLIENT REFERENCE

When we get the call

Two departments publish two different values every month for the same indicator. We trace it back to the sources, settle the definition and rebuild the calculation in a single, dated pipeline. The argument finally moves to the decision instead of the number.

Related disciplines

A file rarely draws on a single discipline.

Development & Product Design

Operable product

Cybersecurity & Audit

Contained exposure
All disciplines in this practice

Planning something in Data & Artificial Intelligence?

Write to us. We respond with an assessment, not a brochure.

Write to us