The six steps, explained
We developed this process over dozens of projects. It is not rigid dogma; we adjust timings and deliverables to fit each client. But the sequence stays the same because skipping a step almost always causes trouble later.
Discovery call
We start with a 45-minute video call. You describe the problem, we ask questions. What data do you have? Where does it live? Who will use the finished system? By the end of this conversation we both know whether the project is a good fit. There is no charge for this call and no obligation to continue.
If the problem turns out to be something a spreadsheet formula or a simple database query can solve, we will tell you. Charging for a machine-learning project when a simpler tool would do the job is not how we want to build a reputation.
Data audit
Before we write a single line of model code, we need to understand your data. We connect to your systems (or you send us anonymised exports) and run a structured audit. How complete are the records? Are there labelling inconsistencies? Is there enough volume to train a reliable model, or do we need to supplement with synthetic data or transfer learning?
This step typically takes five to seven working days. At the end you receive a written report with a traffic-light assessment: green means we can proceed, amber means we need to clean or enrich the data first, red means the data cannot support the objective and we should redefine scope. Roughly one in five audits comes back amber. Red is rare but it happens, and it is far better to know now than after two months of development.
Prototype build
This is where the engineering starts. We select an algorithm family, set up a training pipeline and produce a working prototype. For classification tasks we usually start with gradient-boosted trees because they train fast and give us a quick baseline. If the problem demands it, we move to neural architectures in the next iteration.
The prototype runs on our cloud infrastructure. You can interact with it through a simple web interface or an API endpoint. We schedule a walkthrough session so your team can test it with real inputs and tell us where it gets things wrong. That feedback shapes the next step.
Iteration and validation
We refine the model based on your feedback and our own evaluation metrics. This usually means two to three rounds of retraining with adjusted features, hyperparameters or additional data. We track precision, recall and, where applicable, business-specific KPIs such as cost per misclassification or average processing time.
We share a live metrics dashboard so you can see performance improve between iterations. When the numbers meet the targets we agreed during discovery, we move to deployment planning.
Deployment
Deployment looks different for every client. Some need a containerised API running on their own Azure or AWS account. Others want the model embedded inside an existing application through a REST endpoint. A few prefer a batch process that runs overnight and writes results to a database table.
We handle the infrastructure setup, write deployment documentation and run a handover session with your IT team. If you do not have an IT team, we can manage hosting on your behalf under a monthly support agreement.
Monitoring and support
Models degrade over time as the real world changes. A demand-forecasting model trained on 2023 data will drift if customer behaviour shifts in 2025. We set up automated monitoring that alerts us when prediction accuracy drops below a threshold you define.
Retainer clients get quarterly model reviews where we check performance, retrain if necessary and discuss whether new data sources could improve results. Ad-hoc clients can request a review at any time; we charge by the day for that work.
Tools and technologies we use
Our stack is pragmatic, not fashionable. Python is the backbone: scikit-learn, XGBoost and PyTorch handle most modelling work. For data pipelines we use Apache Airflow or Prefect, depending on client infrastructure. Deployments run in Docker containers, usually on AWS but we have also deployed to Azure, GCP and on-premises servers behind hospital firewalls.
We store experiment tracking in MLflow so every model version, hyperparameter set and evaluation metric is logged and reproducible. When a client asks "why did the model change its recommendation last Tuesday?" we can pull up the exact training run and explain what shifted.
For document processing projects we fine-tune open-source language models rather than relying on third-party APIs. That gives clients full control over their data and avoids ongoing per-token costs that can spiral as document volumes grow.
Questions clients often ask
We hear these regularly during discovery calls. If yours is not listed, get in touch and we will answer it directly.
Most projects reach deployment in eight to fourteen weeks. The data audit takes about a week, prototyping takes two to three weeks, and iteration plus deployment fills the rest. Projects with very messy data or complex integration requirements can stretch to twenty weeks, but we flag that risk early during the audit.
Not necessarily. We can work with anonymised or pseudonymised exports. For clients in healthcare or finance, we often set up a secure environment inside your own cloud account and do all work there. Your data never leaves your infrastructure in that arrangement.
That is what the data audit is for. If the audit shows that the available data cannot support the target accuracy, we will tell you before any development begins. If performance falls short during iteration despite adequate data, we explore alternative approaches: different model architectures, feature engineering or additional data collection. We do not invoice for iteration rounds that fail to improve results.
We price by project phase, not by the hour. A typical engagement including audit, prototype, iteration and deployment runs between £12,000 and £45,000 depending on complexity. The discovery call is free, and the data audit is billed separately at a fixed fee of £1,800 so you can stop after the audit if the findings are not encouraging.
Yes. We have integrated models with Salesforce, SAP, Xero, custom Django applications and several bespoke warehouse management systems. If your system has an API or a database we can query, integration is usually straightforward. Legacy systems without APIs require a bit more creativity, but we have managed it before using file-based interchange and scheduled batch jobs.
Ready to start?
Book a free discovery call and we will tell you honestly whether AI is the right solution for your problem. No slides, no sales pitch, just a technical conversation.
Get in touchPhone: +44 121 702 5722
Email: [email protected]
87 Price Meadow, Birmingham B1 1TF, West Midlands, United Kingdom