AI and machine learning
We are a data and engineering team that builds models where the data supports one. We start by telling you honestly whether it does, and we have not published an AI case study yet, so this page is about method rather than trophies.
Ask for a feasibility read Read what we have and have not done
A feasibility read takes two weeks and produces a yes, a no, or a list of what is missing.
| Is the outcome measurable Can we tell right from wrong | Week 1 |
|---|---|
| Does labelled history exist How much, how clean | Week 1 |
| Would a rule do the job Cheaper, explainable | Week 1 |
| Who acts on the output And are they allowed to | Week 2 |
| Cost of being wrong Decides the whole approach | Week 2 |
Where we stand
Plenty of firms will tell you they have delivered a hundred AI projects. We would rather you could check ours.
We have not
If a vendor with a proven model in your exact domain exists, buy from them. We will say so.
We have
That is the unglamorous 80 percent of a working AI system, and it is where most projects actually stall.
Plain definitions
Both get asked on the first call, and a clear answer to either can end the conversation early in a useful way.
Artificial intelligence is the umbrella word. Machine learning is the part where a system learns a pattern from historical examples instead of being told the rule, which is what powers forecasting, classification and anomaly detection. Generative AI is a narrower branch that produces text or images, and it learned from a corpus that was not yours.
The practical difference is what each one needs from you. Machine learning needs your labelled history. Generative AI needs your documents and a way to check the output. If you have neither, no amount of budget fixes it.
MLOps is the discipline of running a model like software rather than like a science project. Versioned data, reproducible training, a serving path, and monitoring that catches the model drifting away from reality.
You need it the moment a model influences a real decision more than once. A model that was right in March can be quietly wrong by September, because the world moved and nothing told you. Most of what MLOps requires is ordinary engineering, which is the half of this we have running in production.
How we work on this
What people ask us for
These are the enquiries. Our shipped work sits under all of them rather than inside any of them, and the feasibility read is how we tell you which is which.
Our work
This is a data platform, not a model. It is shown here because it is the position most AI projects need to reach before a model is worth training.
Case study, IoT platform
IoT Workz ran device fleets through a SaaS vendor and could not get at their own data. We built a self-hosted platform: telemetry over MQTT, CoAP and HTTP into TimescaleDB, Kafka for decoupling, a rule engine for alarms, and monitoring through Prometheus and Grafana. Their published roadmap moves from there towards predictive maintenance, which is only possible because the history now exists and belongs to them.
Technology
Read this list in two halves. The platform half is running in production today. The model half is what a feasibility read decides you actually need.
Who does the work
An honest feasibility read needs somebody willing to lose the work. That is easier to ask of engineers you know by name than of a sales team paid on signed contracts.
You meet the engineer who writes your feasibility read before it starts. Ask them what would make them recommend a rule instead of a model.
How we work
The models differ in one thing: how much of the management you hand over. Everything else, including who owns the code, is the same in all five.
Changing model later is normal. Augmentation into a dedicated team is the common direction, and the people stay.
Engineers join your team and work inside your process, your board and your code review.
A team that works only on your product, with a lead on our side running delivery.
A long-term engineering site under your standards, where we carry recruitment, HR, payroll and equipment.
04
One price for a scope that is genuinely settled, with the overrun carried by us.
Everything in the centre model, with the whole team moving into your own Vietnamese entity on an agreed date.
FAQs
We have no published case study we can point you at, and we would rather say that than imply a track record we cannot show. What we can show is the data and deployment work underneath such systems, running in production today. If that is not enough for your risk appetite, we understand.
It is a fixed two-week engagement priced as a small project. You get a written conclusion and the data assessment regardless of what you do next.
Then we say no, and tell you what would have to change for the answer to become yes. That happens often and it is a better outcome than a model nobody trusts.
Often, yes, and it is usually cheaper. Training from scratch is rarely the right first move.
Your own cloud tenant, in the region your policy requires. We do not send your data somewhere you cannot see.
You do, both, along with the pipelines and the documentation.
The feasibility read is two weeks at a fixed price and produces a written conclusion whether or not you continue. Getting the data into a usable state is the long pole, because it is a data engineering project wearing a different hat. Modelling and deployment after that is smaller than most people expect. Pricing follows the engagement model above.
It depends on how varied the thing you are predicting is, but the useful question is different. Do you have examples of the outcome, labelled by somebody who knew what they were looking at, and enough of the rare case to matter. Ten thousand readings with four recorded failures will not train a failure predictor, and that is the single most common reason a feasibility read comes back negative.
We are engineers rather than lawyers, so take the classification question to counsel. What we do practically is record where training data came from, keep a person in the decision path where the output carries consequences, document the evaluation so a decision can be explained afterwards, and deploy in your own tenant so nothing leaves your control. Those four hold up under most regimes.
Two weeks later you will have a written answer, including the answer that you do not need a model. That is worth more than a proposal.
