How a typical project unfolds
Every engagement follows the same backbone, though the duration of each phase depends on data complexity and the integration landscape. Here is what to expect.
Discovery call
A 30-minute video call where we listen. We want to understand the business problem, the decision that needs to be made faster or more accurately, and who will use the output. You don't need to prepare slides or a brief; a conversation is enough. By the end we will know whether AI is a reasonable approach or whether a simpler tool would serve you better.
Data audit
This takes one to two weeks. Our data architect connects to your sources, profiles the tables or files, and produces a short report covering volume, completeness, label quality and any privacy concerns. We also document the pipeline: where data originates, how often it updates and where gaps exist. The audit report is yours to keep regardless of whether you proceed with us.
Proposal and scope agreement
Based on the audit we write a fixed-scope proposal: the model type, the target metric (for example, F1 score above 0.82 on held-out test data), the delivery timeline and the total cost. We break cost into two phases so you can approve the prototype before committing to the full build. No hourly billing surprises.
Prototype build
We train an initial model, usually within three weeks. You receive a working prototype with a simple web interface or API endpoint so your team can test it against real inputs. We run a structured feedback session where users flag incorrect predictions, edge cases and usability issues. This feedback directly shapes the next iteration.
Production deployment
Once the prototype meets the agreed accuracy threshold, we package the model for production. Deployment options include a containerised API on your own cloud account, an on-premise server, or our managed hosting. We write monitoring dashboards that track prediction latency, error rates and data-drift indicators so you can see the model's health at a glance.
Ongoing support and retraining
Models degrade over time as the real world shifts. Our support plans include quarterly retraining runs, where we refresh the model on new data and compare performance against the original baseline. If accuracy drops below the agreed floor, we investigate root causes and retrain at no extra charge during the support period.
What happens behind the scenes
During the prototype phase, our engineers experiment with multiple model architectures. For tabular data problems we typically compare gradient-boosted trees (XGBoost or LightGBM) against a shallow neural network. For text tasks we fine-tune a pre-trained transformer, usually a distilled variant to keep inference costs down.
Feature engineering is where most of the value lives. Raw columns rarely predict anything well on their own. We derive rolling averages, interaction terms, time-since-event features and domain-specific ratios that encode business logic the model can't discover unaided. This step often doubles the baseline accuracy.
We version every experiment using MLflow, so you can trace any production prediction back to the exact training run, dataset snapshot and hyperparameter set that produced it. That traceability matters if your sector faces regulatory scrutiny or if you need to explain a decision to a customer.
Frequently asked questions
How much data do we need to get started?
It depends on the problem. For a classification task with five or fewer categories, we can often build a useful prototype from 2,000 labelled examples. Time-series forecasting typically needs at least 18 months of historical records at the granularity you want to predict. If you are unsure, the data audit in step two will give you a clear answer.
Can you work with messy or incomplete data?
Yes. Real business data is always messy. Missing values, inconsistent formats, duplicate records: we deal with these routinely. Part of the audit is quantifying how much cleaning is needed and whether the gaps are random or systematic. Systematic gaps sometimes signal a deeper process issue worth fixing before any modelling begins.
Do we need our own cloud infrastructure?
Not necessarily. We can deploy on your AWS, Azure or GCP account if you have one. If you don't, we offer managed hosting on UK-based servers at a fixed monthly rate. The hosting fee covers compute, storage, SSL certificates and daily backups.
What if the model does not meet the target accuracy?
Our proposal includes a defined accuracy threshold. If the prototype does not reach it after two iteration cycles, you can walk away and pay only for the audit and prototype phase. We absorb the risk of the full build. In practice this has happened twice in 34 engagements, and in both cases the root cause was insufficient training data rather than a modelling failure.
How long does a typical project take end to end?
Most projects run eight to fourteen weeks from discovery call to production deployment. The main variable is the data-preparation phase: clean, well-structured data can shave three weeks off the timeline. Complex integrations with legacy systems tend to add time at the deployment stage.