Common questions about working with us
Most projects fall between £15,000 and £80,000. A single predictive model built on clean, structured data with one data source sits at the lower end. Multi-model systems that involve computer vision, NLP and real-time serving push toward the upper range.
We quote fixed prices per milestone, so you know the total before work starts. The discovery call and data audit (phases one and two) are typically priced together at £2,500 to £5,000. If you decide not to proceed after the audit, that is all you pay.
Eight to sixteen weeks from first call to production deployment. The biggest variable is data readiness. If your data is already in a clean, labelled format, we can move quickly. If it needs significant cleaning, deduplication or manual labelling, that adds two to four weeks.
Smaller single-model projects (for example, a churn prediction model on a well-maintained CRM export) have shipped in as little as four weeks.
Yes. A machine learning model learns from data, so we need access to the data it will train on. Before any transfer happens we sign a mutual NDA and a data processing agreement that specifies exactly what data we receive, how we store it, who can access it and when we delete it.
If your data cannot leave your network (common in healthcare and defence), we can work inside your environment via VPN or a secure virtual desktop. We have done this for three NHS-adjacent projects.
We agree on a performance threshold (accuracy, F1 score, mean absolute error or whatever metric suits the task) during the proof-of-concept phase. If the full model does not meet that threshold, we continue iterating at no extra charge. If after a reasonable number of iterations we cannot reach the target, we refund the milestone fee for that model. This clause is written into every contract.
In practice, 92 percent of our models have met or exceeded the agreed threshold on the first or second iteration. The remaining 8 percent required a scope adjustment, usually switching to a different feature set or redefining the prediction target.
No. Most of our clients have between 50 and 2,000 employees. We have also worked with startups as small as 12 people and with divisions of FTSE 250 firms. The deciding factor is not company size but whether you have enough data to train a useful model. A startup with 50,000 transaction records can be a great fit. A large enterprise with fragmented, unlabelled data might need a data collection phase first.
Yes. We deploy on AWS, Azure and GCP. If you already have a cloud account with specific networking rules, IAM policies or compliance requirements, we work within them. We also deploy on-premise for clients who need it, though on-premise deployments typically add one to two weeks to the integration phase because of hardware provisioning and firewall configuration.
Every project includes three months of post-launch support. During that period we monitor model performance, fix bugs and handle one retraining cycle if the data distribution shifts. After three months you can extend support on a monthly retainer (typically £1,500 to £3,000 per month depending on complexity) or take over operations internally.
We write detailed runbooks and architecture documentation during the deployment phase so your team can maintain the system independently. If you have data engineers or ML engineers in-house, we run a handoff workshop to walk them through the codebase.
We follow UK GDPR for all personal data processing. For sector-specific regulations (FCA rules for financial services, NHS DSPT for health data, MOD security clearance for defence), we tailor our data handling procedures to the relevant framework before any data is transferred.
All data in transit uses TLS 1.3 encryption. Data at rest is encrypted with AES-256. Access is restricted to the specific engineers assigned to your project, and we rotate credentials every 90 days. We can also work with anonymised or pseudonymised datasets if full personal data is not needed for the model.
In our experience, no. The models we build automate specific, repetitive tasks: sorting emails, flagging anomalies, generating demand forecasts. The people who used to do those tasks manually end up spending their time on work that requires judgement, context and customer relationships. One logistics client reassigned two full-time admin staff to client management roles after we automated their shipment classification process. Both employees stayed with the company.
Technical questions
What programming languages do you use?
Python for all machine learning and data engineering work. SQL for data transformations. Bash and Terraform for infrastructure automation. If your application layer is in Java, Node.js or .NET, we write the integration glue in that language so your developers can maintain it.
Can you fine-tune large language models?
Yes. We fine-tune open-weight models (Llama, Mistral, Phi) on your domain-specific text using LoRA or QLoRA to keep compute costs reasonable. A typical fine-tuning run on a 7-billion-parameter model costs between £200 and £800 in cloud GPU time. We do not fine-tune proprietary models like GPT-4, but we can build retrieval-augmented generation (RAG) systems that use them via API.
How much data do we need?
It depends on the task. For tabular prediction (churn, demand, pricing), a few thousand rows with clean labels is often enough to build a useful model. Computer vision needs at least 500 labelled images per class for fine-tuning a pre-trained backbone. NLP classification works well with 1,000 or more labelled examples per category. We assess this during the data audit and tell you honestly if you need more.
Still have a question?
We are happy to answer anything that is not covered here. The fastest way to reach us is email. We typically reply within one working day.
Email: [email protected]
Phone: +44 113 218 4465
Address: United Kingdom, Scotland, North Streich-Skilesworth, UZ39 3GD, 2 Baker Street