Machine learning development services for businesses in the US, UK, Canada and Australia that want one decision made faster: a ranking, a forecast or a flag on records that look wrong. We pilot it, measure it before and after, and you keep the code, the model and the pipeline in your own cloud account.
BioBuild material recommendation engine, live since January 2025 · 4.5 of 5 in a verified Clutch review
WHAT IT IS
Machine learning development services build models that learn from your historical data to make one business decision faster or better: ranking options, predicting a number, or flagging records that look wrong. The work covers scoping, data preparation, training, integration into your software and retraining. Our shipped example is a material recommendation engine for BioBuild in Sweden.
It is not generative AI. A model predicts a number or a ranking from your own history; generative AI writes text or answers from documents. They need different data, testing and budgets.
Put the best few choices in front of a person, in order, with the reason.
Materials, products, suppliers, next actionsEstimate demand, cost, time or risk from your history, with the likely error shown beside it.
Stock to order, delivery dates, project costSort records into groups and pick out the ones that do not fit, before someone acts on them.
Payments, claims, sensor readings, data entryWHEN TO CALL US
If two or more of these sound like your business, one decision is probably worth a feasibility study.
A decision we make every week takes days of spreadsheet work.
The answer is out of date by the time it is ready.
Our forecasts are guesses, and we know what being wrong costs.
That cost of being wrong is the ceiling on what a model is worth.
A vendor says AI, and nobody can tell us if our data is enough.
You need a yes or no on feasibility before a build quote, not after it.
Someone built a model once, and nobody has retrained it.
It was right on launch day, and nobody has measured it since.
Our product needs ranking or recommendations.
Search results or a shortlist that should get better with use.
Our experts are the bottleneck, and they are hard to hire.
A model can do the first pass, so the expert reviews a shortlist.
Machine learning fits when you need a number or a ranking learned from history. For many problems something simpler works better, and costs less to run.
| If you | What usually fits | Example |
|---|---|---|
| Can write the rule down | Rules in code, no machine learning | Tax and approval rules, routing by fixed criteria |
| Need text read from documents, speech or images | A prebuilt service | Invoice agent on Azure Document Intelligence |
| Need answers or drafts from your documents | Generative AI | AI chatbots and generative AI |
| Need a number or a ranking learned from history | Machine learning | BioBuild material ranking |
| Have data spread across spreadsheets and systems | Data engineering first | Data engineering consulting |
| Are not sure it will pay | A feasibility study | The scoping steps below, with a go or no-go at the end |
WHAT WE DELIVER
Six kinds of work. Each card says whether a published Metwaves case study backs it.
Models that rank options for a person to choose from, such as materials, products or next actions, with the trade-off shown.
BioBuild material engine →Demand, cost, time or risk predicted from your history, with the error range shown next to the number.
Sorting records into categories and flagging the ones that look wrong before a person acts on them.
Prebuilt services first, such as Azure Document Intelligence, and custom training only where they fall short.
Invoice agent →One decision, one metric and your own data, ending in a written go or no-go before you fund a full build.
How BioBuild was scoped →The model served as an API, watched for drift and retrained on a schedule.
BioBuild, retrained quarterly →What you keep in every case: the code, the trained model, the training pipeline and the data, in your own cloud account.
SCOPING
Pick one decision, one metric and one pilot, and agree what good enough means before anyone trains a model. These are the five steps we used to scope BioBuild.
One frequent decision, such as which supplier, price or material.
Time, cost or error rate today, measured the way you will measure the model.
What history you have, where it lives and whether outcomes were recorded.
The result that justifies a production build, written down before any modelling starts.
One project or one team, with experts reviewing every output. Then go, change course or stop.
DATA READINESS
Most first machine learning projects are data projects. Tick what is already true for you; the panel says where to start.
A model is one box in a longer chain. Data is shaped, the model is trained and checked against the agreed metric, served to your software, and retrained on the outcomes it records.
MLOPS
A model is accurate on the day it ships; your data drifts after that. These six practices run around every model we keep live.
Alert when new data stops looking like the training data: new suppliers, odd prices, missing fields.
Compare predictions with what happened, using the metric agreed at scoping.
BioBuild writes back actual cost and material performance after each project and retrains every quarter.
Every model version is kept with the data it was trained on, so a worse model can be replaced by the last good one.
Experts can overrule a recommendation, and their corrections become training data.
Hosting, retraining and monitoring are reported each month, separate from the build.
Six steps, each ending in something you can check. BioBuild went from the start of work in August 2024 to go-live in January 2025, pilot included.
The decision, the metric, the baseline and a data check. Ends in go or no-go.
Pipelines from your systems, and labels built where none exist.
One project or team, every output reviewed, every correction kept. BioBuild's pilot was running two months after work began.
An API and a dashboard or screen beside the tool your team already uses.
Drift checks, scheduled retraining and a monthly report.
Code, model and pipeline documented in your account. Keep us for retraining, or run it yourselves.
CASE STUDY
BioBuild, a Swedish construction firm, chose materials by hand in spreadsheets, checking prices, structural ratings and carbon data separately. Every design change restarted the loop.
We projected its IFC models into Neo4j, trained Random Forests and k-means on its past projects, and used a genetic algorithm to balance cost, carbon and structural fit, with the weighting set by the firm. A ranked shortlist updates beside the modelling tool whenever the design changes, and actual cost and material performance are written back after each project for the quarterly retrain.
One decision per use case. Only the first row is Metwaves work; the others show the same method on other decisions.
| Industry | Decision | Model type | Status |
|---|---|---|---|
| Construction | Which material to use per element | Recommendation: Random Forests, k-means, genetic algorithm | Shipped: BioBuild |
| Healthcare | Which appointments are likely to be missed | Classification | Example, not a Metwaves project |
| Fintech | Which transactions need a second look | Anomaly detection | Example, not a Metwaves project |
| Logistics | When a delivery will arrive | Prediction | Example, not a Metwaves project |
| Retail and eCommerce | How much stock to order | Forecasting | Example, not a Metwaves project |
COST
You pay by phase and can stop after any of them. The monthly running cost is shown apart from the build.
We do not publish prices: data work drives most of the cost and differs for every firm.
The feasibility study is a fixed quote. What follows depends on what the study finds.

The feasibility study, the pilot and each production phase are quoted once the one before has answered its question.

For model support and retraining after go-live, billed monthly for the work done.

Our engineers working inside your team, billed monthly, when you run several models and want the work close at hand.

Time and material while the data is explored, then a fixed price once the build is clear.
The BioBuild engine also uses Neo4j, BIMserver and IFC models, and the invoice agent uses Azure Document Intelligence. We add tools such as PyTorch or scikit-learn when a project needs them, and say so in the proposal.

Ask these five questions of anyone you shortlist, us included. The last one is where we lose.
| Question to ask | A good answer | A warning sign |
|---|---|---|
| Can we see a model in production, and how is it retrained? | A live system with a retraining schedule. | A list of algorithms and a demo notebook. |
| How did you measure the result? | A baseline taken before go-live and the same measure after it. | Accuracy on a test set, and nothing about the business. |
| Who owns the model, the code and the training data? | You do, in your cloud account, with the pipeline. | The model is hosted on their platform, under their account. |
| Will you tell us when machine learning is not needed? | Yes, and point you to a rule or a prebuilt service. | Every problem gets a custom model. |
| How many custom ML systems have you published? | A number you can check, with each one described. | Dozens of logos and no detail. We publish one, in full; if you need many references, we are not that firm. |
Machine learning rarely stands alone: the data has to arrive, and the result has to land somewhere.
QUESTIONS & ANSWERS
Straight answers on cost, time, data, ownership and what happens after launch. If you don't see an answer to your question, you can send us a message from our contact form.
It is priced by phase: feasibility study, pilot, production build, then a monthly running cost. Data work, labels, integration depth and the accuracy target move the price most. Each phase is quoted separately, so you can stop at any point.
BioBuild went from the start of work in August 2024 to go-live in January 2025, including a pilot on one live project. A feasibility study is much shorter, because it ends in a decision, not a system.
Enough history for the one decision you want to improve, plus a record of outcomes. Coverage matters more than volume. BioBuild had no labelled dataset; we built one from its past projects.
Machine learning predicts or ranks from your historical data. Generative AI writes text or answers questions from documents. See our generative AI page for the second.
It is monitored for drift and retrained on recorded outcomes. The BioBuild model is retrained every quarter.
You do: the code, the trained model, the training pipeline and the data, in your own cloud account.
The pilot has a threshold agreed before modelling starts. If the model misses it, you stop with a written finding on why, not a half-built system.
Discover IT Industry Insights and Technology Trends
Support for .NET 8 and .NET 9 ended on November 10, 2026. What that means for apps still on them, how risky it is, and how to plan the move to .NET 10.
Learn More >Nine AI agent development firms a small or mid-sized business can shortlist, with Clutch figures, minimum budgets and the weaknesses their own clients name. Published by Metwaves, which is not ranked.
Learn More >An architecture review of the four big enterprise AI agent platforms: where each one reasons, acts, finds its data and connects to outside agents, what it costs, and when a custom build is the better choice.
Learn More >
Let us help you with free analysis & estimate.
Our product expert will contact you within 12 hours.