How our Artificial Intelligence projects work
Every engagement follows six stages. The whole sequence typically takes six to twelve weeks, depending on data complexity and the number of model iterations your use case requires.
Six stages from first call to production
Below is the actual workflow we follow. Timelines are averages; some steps overlap, and simpler projects skip certain sub-tasks entirely.
Discovery call and data audit
We start with a 45-minute video call. You describe the business problem; we ask about the data you already collect, where it lives, and how clean it is. After the call we sign a mutual NDA and request sample exports from your databases, spreadsheets, or APIs.
Within five working days we deliver a short data-audit report. It covers volume, feature completeness, label availability, and any obvious quality issues like duplicates or missing timestamps. This step is free for all prospective clients.
Problem framing and success criteria
Machine learning can answer many questions, but only if you phrase them precisely. We translate your business goal into a measurable ML objective. For example, "reduce customer churn" becomes "predict which accounts will cancel within 90 days with at least 78% recall at 60% precision."
We also agree on a kill threshold. If the proof of concept cannot beat a simple baseline by a meaningful margin, we stop and you pay only for data preparation. This protects you from spending £15,000 on a model that adds no value over a spreadsheet rule.
Data preparation
Raw data is rarely model-ready. Our data engineer writes extraction scripts, handles joins across your systems, fills or flags missing values, and engineers features. For NLP tasks, this includes tokenisation, language detection, and annotation of a gold-standard evaluation set.
We version every dataset with DVC and store artefacts in your cloud bucket. That means you can reproduce any experiment months later without guessing which CSV was used.
This phase usually takes one to three weeks. Complex multi-source integrations with privacy constraints (healthcare, finance) sit at the longer end.
Model development and proof of concept
We train candidate models, compare them on held-out test data, and present results in a Jupyter notebook you can run yourself. For tabular data we typically benchmark gradient-boosted trees against a neural baseline. For vision and language tasks we fine-tune pre-trained transformers.
You see confusion matrices, calibration plots, and feature-importance charts. We walk you through what the model gets right, where it fails, and why. If accuracy meets the agreed threshold, we proceed. If not, we discuss whether more data, different features, or a revised objective could close the gap.
Deployment and integration
A model sitting in a notebook helps nobody. We package the chosen model as a REST API (FastAPI on Docker) or embed it directly into your existing application via a Python or Node SDK. Inference latency targets are agreed upfront; most of our deployments respond in under 200 ms at the 95th percentile.
We configure CI/CD pipelines so that code changes trigger automated tests and staged rollouts. The deployment runs on your own AWS, Azure, or GCP account. We never host client models on infrastructure we control, because that creates a dependency you should not have to accept.
Monitoring and iteration
After launch we set up drift-detection alerts that fire when incoming data distributions shift beyond a configurable threshold. A Grafana dashboard shows prediction volume, latency, and accuracy metrics in real time.
On our retainer plan (from £1,200 per month) we review these dashboards weekly, retrain models quarterly or when drift triggers, and respond to critical issues within four hours. Clients who prefer to manage models in-house receive a two-day handover workshop covering monitoring, retraining, and rollback procedures.
What to expect at each stage
Your time commitment
During discovery and framing (stages 1 and 2) we need roughly three hours of your team's time across two meetings. After that, involvement drops to a 30-minute weekly sync and occasional Slack questions about domain context.
If your data lives in systems that require IT access requests, factor in a few days of lead time. We have seen projects delayed by two weeks simply because a firewall rule needed a change-management ticket.
Deliverables you own
At the end of every engagement you receive: the trained model weights, all preprocessing code, a reproducible training pipeline, API source code, deployment configuration files, and written documentation covering architecture decisions, known limitations, and maintenance instructions.
Everything is committed to a Git repository under your organisation's account. There is no vendor lock-in, no proprietary runtime, and no recurring licence fee for the software we write.
Questions we hear often
Most engagements fall between £8,000 and £35,000, depending on data complexity and the number of models required. A straightforward tabular-data classifier with clean, labelled data sits at the lower end. Multi-modal systems that combine text and image inputs cost more because they need separate preprocessing pipelines and longer training runs. We always quote a fixed price after the free data audit, so there are no surprises.
No. Many of our clients have no ML expertise internally. That is exactly why we include thorough documentation and a handover workshop. If your developers can write Python and manage Docker containers, they can maintain a deployed model. For teams without any technical staff, our retainer plan covers all ongoing maintenance.
Our kill clause activates. You pay only for the data-preparation work completed up to that point, and you keep all code and cleaned datasets. We will explain why the target was not met and whether additional data collection, different labelling, or a revised problem frame could change the outcome. About 15% of our proof-of-concept projects trigger this clause, usually because the underlying signal in the data is too weak for the desired precision.
All processing happens on your cloud infrastructure. We request temporary access credentials scoped to the specific resources we need, and we revoke them at project close. Sample data shared during the audit phase is stored on an encrypted drive and deleted within 14 days of the audit report delivery. We are happy to sign a Data Processing Agreement if your compliance team requires one.
Yes. Two of our current clients run models on local GPU servers for regulatory reasons. We configure the same Docker-based deployment pipeline but target your on-premise hardware instead of a cloud instance. The only requirement is SSH access during setup and a machine with at least 16 GB of RAM and a CUDA-capable GPU for inference-heavy workloads.
Ready to start?
Book a free discovery call and data audit. We will tell you honestly whether AI fits your problem and what it will cost.
Get in touch