How we deliver Artificial Intelligence projects

Every engagement follows the same five-phase structure. Timelines vary by complexity, but the sequence stays consistent because it works.

Five phases, start to finish

Below is the path from first conversation to a model running in your production environment, with realistic durations for each phase.

1. Discovery and data audit

Duration: 1 to 2 weeks

We start with a half-day workshop at your office or over video. The goal is blunt: figure out whether the problem you want to solve is actually a good fit for machine learning, or whether a simpler rule-based approach would serve you better. We have talked clients out of AI projects before, and we would rather do that than build something that gathers dust.

During this phase we also get read access to your data sources. Our engineers profile the tables, check for missing values, measure class balance and estimate whether you have enough labelled examples to train a useful model. You receive a written data-readiness report at the end, regardless of whether we proceed.

2. Problem framing and proposal

Duration: 3 to 5 days

Once we understand the data, we write a formal proposal. It specifies the model type (for instance, gradient-boosted trees for tabular churn prediction, or a fine-tuned transformer for contract-clause extraction), the success metric we will optimise (F1 score, mean absolute error, precision at a given recall threshold), and the acceptance criterion that determines whether the model is good enough to deploy.

The proposal also lists what we need from your side: a point of contact who can answer domain questions, access to a staging environment, and any compliance constraints around data residency. We attach a fixed price and a delivery date. You sign it or ask for changes; nothing moves forward without written agreement.

3. Data preparation and feature engineering

Duration: 2 to 4 weeks

This is the phase most people underestimate. Raw data rarely arrives in a shape that a model can learn from. We clean, join and transform it, then build feature pipelines that can run automatically when new data arrives. If labelling is needed, we set up an annotation workflow and quality-check every batch.

All pipeline code is version-controlled in a Git repository that belongs to your organisation. We do not keep a separate copy. If you ever want to bring the work in-house or hire another vendor, the code is yours.

4. Model training and validation

Duration: 2 to 6 weeks

We train candidate models, compare them on a held-out test set, and run error analysis to understand where each one fails. A model that is 94% accurate overall but misses a critical edge case is worse than one that is 91% accurate with evenly distributed errors, so we dig into the confusion matrix rather than quoting a single headline number.

You see results as they come in. We share a live dashboard where you can filter predictions by date, category or confidence score. If the numbers are not meeting the acceptance criterion, we discuss whether to collect more data, try a different architecture, or adjust the target metric. Honest feedback loops save everyone time.

5. Deployment, monitoring and handover

Duration: 1 to 3 weeks

We package the model as a containerised REST API or, for edge deployments, as an optimised binary. The service includes health checks, request logging and an alerting layer that notifies your ops team if prediction latency or error rate exceeds agreed thresholds.

After deployment we run a two-week burn-in period during which we monitor performance against live data. Drift detection compares the distribution of incoming features to the training set and flags significant shifts. Once burn-in passes, we hand over documentation, run a knowledge-transfer session with your engineering team, and close the project.

Optional maintenance contracts cover quarterly model retraining, feature-store updates and on-call support. About a third of our clients choose this option.

What you get at the end

When a project wraps up, you own every artefact. That includes the trained model weights, the feature-engineering pipeline, the deployment container, the monitoring dashboard configuration and a written technical guide that explains how each component works and how to retrain the model on fresh data.

We also provide a non-technical summary for stakeholders who need to understand what the model does and where its limitations lie. This document covers the training data period, known failure modes, and the conditions under which the model should be retrained or replaced.

There is no lock-in. If your internal team wants to take over maintenance after handover, we support that. If you prefer us to keep running retraining cycles, we offer 6- and 12-month maintenance agreements at a flat monthly rate.

Model performance dashboard on laptop screen

Questions we hear often

If your question is not here, email us at [email protected] or call +44 161 446 1556.

Most of our projects fall between £15,000 and £80,000. A straightforward tabular prediction model with clean data sits at the lower end. A multi-modal system that combines NLP and vision, or a project that requires significant data labelling, sits higher. We always quote a fixed price before work begins.

Not during the project. We handle everything from data engineering to deployment. After handover, you will need someone who can monitor the dashboard and trigger a retraining run when drift alerts fire. That can be a DevOps engineer; it does not have to be a data scientist. Alternatively, our maintenance contract covers it.

On your infrastructure. We deploy to AWS, Azure or GCP, whichever you already use. For edge use cases (factory floor cameras, for example) we deploy to NVIDIA Jetson or similar devices on your premises. Your data never touches our servers.

We discuss the gap honestly. Sometimes collecting more training data closes it; sometimes the problem is harder than either side expected. If we cannot reach the agreed metric after a reasonable effort, you pay only for the work completed up to that point. We have had to invoke this clause twice in four years.

Yes. We have completed projects involving NHS patient administration data and FCA-regulated financial records. We sign data-processing agreements, work inside your VPN, and our engineers hold current DBS checks. If your compliance team needs to audit our practices before we start, we welcome that.