Skip to main content

Dev Station Technology

AI and machine learning

Most AI projects stall on the data

We are a data and engineering team that builds models where the data supports one, and we say early whether it does. We have not published an AI case study yet, so this page describes method.

What a feasibility read checks

Is the outcome measurableCan we tell right from wrong
Does labelled history existHow much, and how clean
Would a rule do the jobCheaper, and explainable
Who acts on the outputAnd are they allowed to
Cost of being wrongDecides the approach

Half of these reads end with advice not to build a model.

2 weeksA feasibility read, with a written yes, no or list of gaps
0AI case studies published, and we say so
YearsOf queryable readings on the IoT Workz platform
BaselineBuilt before any model, so you know what it must beat
What an AI system is made of

The data half is shipped, the model half is decided by the read

From history to a model you can watch
Data
Ingestion and labellingUsually the longest step
Time series historyTimescaleDB and PostgreSQL
Model
Baseline firstA rule or a simple statistic
Python logo
Python modelsscikit-learn, XGBoost, LightGBM
Run
DeploymentIn your own infrastructure
Monitoring and drift alertsSame machinery as pipeline monitoring

Solid boxes run in production today. Dashed boxes are what a feasibility read decides you need.

Fit check

Is a model the right move?

Pick us when

  • The outcome is measurable
  • Labelled history exists, or can be built
  • A rule would not do the job
  • Someone is allowed to act on the output

Look elsewhere when

  • A vendor with a proven model in your exact domain exists
  • You need novel research architectures
  • No history exists at all
  • You want accuracy figures from a past project that we can quote
How we work on this

Five steps, and we stop at any of them if the answer is no

From feasibility to a watched model
2 weeks

Feasibility read

Measurable outcome, history, rule or model, who acts, cost of error. You get a written answer either way.

Next

Make the data usable

Ingestion, labelling, storage and reconciliation. Our strongest ground.

Then

Baseline first

A rule or statistic the model has to beat. Sometimes it is good enough.

After

Model and evaluate

Established libraries, tested on data the model has never seen.

Finally

Deploy and watch

Serving and drift alerts, because a model right in March can be wrong by September.

Predictive maintenance needs years of readings and labelled failures. Most companies have the first and not the second.

Technologies

What this would be built on

Established libraries first. PyTorch or TensorFlow only where the problem needs them.

Builds

The foundation a model needs

IoT Core Engine self hosted platform built by Dev Station
Case study, IoT platform

Years of readings, owned by the client, ready to be learned from

IoT Workz could not reach its own data through a SaaS vendor. We built a self-hosted platform with telemetry over MQTT, CoAP and HTTP into TimescaleDB, Kafka for decoupling and monitoring through Prometheus and Grafana. Predictive maintenance is only possible next because the history now exists and belongs to them.

YearsQueryable history for future models

Ask for a feasibility read

Three fields. An engineer reads it, and you get an answer within one working day.

  • Two weeks, fixed scope, written conclusion
  • The answer may be that a rule beats a model
  • No obligation, and no phone number needed to start
FAQs

Questions we get on the first call

AI, machine learning, generative AI: what is the difference?

Artificial intelligence is the umbrella word. Machine learning learns a pattern from historical examples and powers forecasting, classification and anomaly detection. Generative AI produces text or images. Machine learning needs your labelled history, and generative AI needs your documents and a way to check the output.

What is MLOps, and when do you need it?

MLOps is running a model like software: versioned data, reproducible training, a serving path and drift monitoring. You need it the moment a model influences a real decision more than once.

Do you have an AI case study?

No, and we say so. We have built the data platform that models depend on, and the IoT Workz platform runs in production.

What would predictive maintenance need?

Years of sensor readings and labelled failures. Most companies have the readings and not the labels, which is why a feasibility read comes first.

Tell us the decision you want made automatically

A feasibility read takes two weeks and gives you a yes, a no, or a list of what is missing.

Let's Talk