UNDAMENTAL. From big data to rich data.
Big Data is precognitized. Designed by humans, for humans. Everything we capture gets flattened to fit the boundaries of our own minds. Those terabytes are heavily biased and low in information.
Rich Data allows cheaper and higher complexity streams capable of describing not just the bits we deem important, but the whole of reality.
Before intelligence can learn from the world, we've already flattened it.
Most instruments are built to report one thing and suppress the rest as interference. That interference is usually other things happening: impacts, flows, footsteps, the structure itself.
Noise is the word we use for signal we perceive but cannot understand.
Calibration is a formal decision about what to subtract. Often, you are subtracting the most valuable parts.
Machine learning reads noise like a painter sees colour. The intelligence of tomorrow possesses intuition that exceeds the imagination.
Rich Data expands what we can infer from the world, without requiring us to record all of it.
Sensor geometry, placement and recording are design choices. The overlaps between measurements narrow down what is physically happening in ways no single sensor can.
Prediction depends on what was observed.
A data lake keeps what it receives. It cannot keep what was never measured. That decision is made upstream, at the sensor, and that is where we work.
Storage is the cost that shows up on an invoice. The larger one is not seeing.
Existing sensors can be cut to what they are good at, and the budget that frees pays for sensors placed to see the rest. Total storage stays comparable to today's.
The Romans could not have guessed what data we would want from them.
This is the first time it has been possible to collect for future needs with some confidence, because we can already see the kind of reader those needs will have.
What a single vibration sensor records, opened out into the processes that made it.
Method · Illustration How sources that share one channel can be told apart.
Different processes leave different shapes: a shaft is periodic, a resonance rings and decays after impacts, footsteps are sparse, turbulence is broadband, drift is slow. Those differences are what a decomposition uses.
Where two sources share the same signature, one channel cannot separate them. That limit is why overlapping measurements matter (Fig. 3).
Much of what we file under noise is a pattern spread wider than we can follow.
Method · Illustration Why a pattern can be invisible in every signal we look at, and plain across all of them.
A person can follow a handful of signals at once, and judges the rest by eye. Calibration, filters and dashboards are built to that measure: they keep what a person can read and call the remainder noise.
Machines are not held to that measure. They read every channel together and keep the relationships between them. Much of what we have called noise was never too small to see, only too wide. The limit was ours.
Records made for one purpose, and the questions they answered centuries or decades later.
Evidence · Cited Five records, each read for something its makers could not have intended.
Existing methods already use the signal flat pipelines discard: anomaly detection, predictive maintenance, quality inspection. Some deploy in weeks.
Your own AI systems work through your proprietary data banks, finding patterns, building predictive models, surfacing hidden risks and opening new services.
Depth in time accumulates. Each new generation of models draws more from the same archive, and no competitor can buy back the years you recorded.
The same hardware, built into machines or placed to overlap, and what each arrangement can tell apart.
Method · Illustration How an overlapping arrangement separates and locates sources.
How signals travel, attenuate and mix through a structure, giving the timing and strength differences that separate and locate sources.
What is known about each signal's form: periodic, sparse, slowly varying, bounded. They let more be recovered than the raw readings seem to hold.
Machine learning for the parts no equation describes, trained and checked against the physics rather than in place of it.
One machine, forecast from what was observed up to now.
Method · Illustration Why a model needs more of the system observed than a dashboard does.
How the same storage budget pays for diagnosis, discovery and a live model.
Method What each job needs, and how storage is scheduled to serve all three.
Completeness around events, so failures can be explained with confidence.
Most of what is worth acting on recurs, so it needs enough events, seen without bias, rather than all of them. Rare one-offs are a trade-off, covered by the buffer.
Computed from everything measured; its outputs are kept continuously.
Evidence · Simulation Published methods, run live on simulated signals. Nothing here is a deployment result.
From magnifying available data to building state-of-the-art sensing.
Our multidisciplinary team works alongside your engineers. Our process starts inside your operational reality and takes about six months from first visit to a running pilot.
On site with your engineers. We map what each system senses, at what rate, what it keeps and what it discards.
Deliverable An inventory of discarded signal, what it would cost to retain, and which of it already has known machine learning uses.
Modality, placement, sampling and retention design for the chosen scope, built to run beside what is already there.
Deliverable A pipeline specification with a storage and cost model.
Deployment on a contained scope, alongside existing reporting.
Deliverable Rich capture running in production, and a first account of what it shows.
Founder and CEO
Guilhem founded UNDAMENTAL on the view that many of the limits placed on machine intelligence are set upstream, in how the world is measured and represented. He remains personally involved in every client engagement, from initial diagnostic through to implementation, so that each is shaped around the organisation it serves.
The breadth of his career, which began in the arts and now combines scientific depth with pioneering innovation, underpins UNDAMENTAL's approach: diagnostics and implementation plans developed from first principles, rather than inherited from the conventions of a single discipline.
Chief Marketing Officer
Julio leads commercial strategy, positioning and partnerships. He brings luxury-sector excellence to client stewardship, ensuring each pilot is positioned to withstand scrutiny within the client organisation and that its value is articulated with equal clarity from operational teams to the board.
He has spent more than a decade driving commercial growth at international scale, leading sales, marketing and partnership functions for global hospitality and luxury groups, technology companies and founder-led ventures across Europe, the Americas and Asia, and works in four languages.
Tell us about your project.
UNDAMENTAL