Sciematics
Talk to a scientist
Data science and artificial intelligence

Models that survive contact with reality.

Sciematics builds machine learning systems that are measured honestly, monitored after launch, and abandoned when the evidence says they should be. Most of what is sold as AI would not pass that test.

2024
Founded
0
Models in production
0
Projects we advised against
Position

The uncomfortable first question

Almost every client arrives certain they need a model. A large share of them do not. Their data has leakage, or too few labels, or the decision it would inform is already made by a rule that works.

So we ask whether the data can support a model before we ask what the model should do. Nine times we have answered no and said so. That number is on this page deliberately.

0
Models in production
100%
Ship with drift monitoring
100%
Weights owned by clients
0
Metrics quoted without an interval
Why Sciematics

The field has an evaluation problem

A model that reports 94 percent accuracy on a slide has told you almost nothing. On what test set. Against what baseline. With what uncertainty. Measured when.

Those four questions eliminate most of what is presented as artificial intelligence in a commercial setting. A classifier that beats a coin flip is impressive to nobody, and a classifier evaluated on data it was trained on is not evaluated at all.

We publish confidence intervals, name our baselines, and hold out a test set that no one sees until the end. This is ordinary scientific practice. It is unusual enough in industry that it has become our position.

Read the method
Capabilities

Eight disciplines

Data Readiness Assessment

Before any model, an honest audit of whether your data can carry one.

Predictive Modelling and Forecasting

Demand, churn, risk, and price. Models judged against a baseline you agree beforehand.

Language and Document Systems

Retrieval, extraction, and assistants over your own documents. Rarely a fine tune, almost never a new model.

Computer Vision

Detection, segmentation, and quality inspection on real imagery, with the lighting you actually have.

Machine Learning Engineering

Taking a notebook that works once and making it a system that works every day.

Data Engineering and Platform

Warehouses, pipelines, and the unglamorous foundation everything above it depends on.

Analytics and Decision Support

Sometimes the answer is a query and a chart. We will say so.

AI Governance and Assurance

Independent evaluation of models, including ones we did not build.

Responsible AI

A model that affects a person deserves scrutiny

If a system decides who gets a loan, a diagnosis, or a job interview, then accuracy is not the only number that matters. Who does it fail, and how often, and can a human overturn it.

We evaluate across subgroups, document known limitations in a model card, and design the human oversight before deployment rather than after an incident.

Our governance practice

Four commitments we can be held to

BaselineNamed before modelling starts
Test setHeld out, opened once
UncertaintyReported on every metric
MonitoringShipped with the model
Questions

What clients ask first

Usually whether their data is good enough. It is the right question, and it is rarely asked early enough.

Ask us directly

That is the first question we answer, and often the answer is not yet. Before any modelling we run a data readiness assessment covering volume, label quality, leakage, and drift. If the honest answer is that you need six months of better instrumentation first, you will hear that instead of a proposal.

Almost never from scratch. Most business problems are solved by retrieval, careful prompting, and a small fine tune on a well curated dataset. Training a foundation model is a research budget, not a business decision.

With a held out test set the model has never seen, a baseline it must beat, and a metric agreed before we start. Every number we publish carries a confidence interval, because a point estimate with no uncertainty is a marketing claim, not a result.

It will. Data drifts and the world moves. Every model we ship comes with drift monitoring, an alert threshold, and a retraining plan. A model without monitoring is a liability with a launch party.

You do. Weights, code, notebooks, and pipelines are yours from the first commit. We do not train on your data for other clients, and we do not retain a copy after the engagement ends.

Frequently. A large share of problems presented to us as machine learning problems are solved better by a SQL query, a rules engine, or fixing the process that generates the data. Saying so costs us revenue and is the reason clients return.
Next step

Send us your data problem. We will tell you whether it is one.

A working data scientist reads every enquiry. You will get an honest read on whether your data can support the model you have in mind, before anyone quotes you a number.