We developed this process over 40+ projects. It keeps timelines predictable and gives you clear decision points along the way.
1. Discovery and data audit
We spend the first week getting to know your business, your data and your goals. This starts with a 90-minute kickoff call where we ask about the decision you want the model to support, the systems that hold the relevant data and the people who will use the output. After the call we request read-only access to sample data and run a set of checks: volume, completeness, label distribution, class imbalance and potential bias. By the end of the week you receive a scope document that describes the proposed solution, the expected accuracy range, the integration plan and a binding fixed-price quote. The discovery fee is £1,200, payable upfront.
2. Data preparation
Raw data is rarely model-ready. This phase covers cleaning, deduplication, feature engineering and, where necessary, annotation. For NLP projects we often build custom tokenisation rules or domain-specific stop-word lists. For vision projects we standardise image sizes, apply augmentation (rotation, crop, colour jitter) and split the dataset into train, validation and test partitions. If your data lives in multiple systems, we write ETL scripts to pull it into a unified format. This phase typically takes one to two weeks.
3. Model development
We train candidate models, compare architectures and tune hyperparameters using cross-validation. You see interim results in a shared notebook (Jupyter or Colab) so you can ask questions and steer priorities. For tabular data we usually start with gradient-boosted trees and move to neural networks only if the data warrants it. For text and image tasks we fine-tune pre-trained transformers or convolutional networks. This phase runs two to four weeks depending on iteration cycles.
4. Validation and testing
Before anything goes near production we run the final model against a held-out test set that it has never seen. We report precision, recall, F1 and, where relevant, calibration curves and fairness metrics. If the results fall below the threshold agreed in the scope document, we go back to step 3. We also run stress tests to check latency under load: for real-time APIs we target p95 response times below 200 ms.
5. Deployment
We package the model as a Docker container and deploy it behind a load balancer on the cloud provider of your choice (or on-premises if required). The deployment includes a health-check endpoint, structured logging and a Grafana dashboard that tracks prediction volume, latency and drift indicators. We write integration glue so your application can call the model via REST or gRPC. If you use a CI/CD pipeline, we add the model build to it so future retraining flows through the same process as the rest of your software.
6. Handover and support
We deliver documentation that covers the model architecture, the training data lineage, the retraining procedure and the monitoring alerts. A two-hour walkthrough session with your engineering team ensures they can operate and retrain the model independently. For 90 days after launch we monitor the deployment and fix any issues at no additional cost. If accuracy drifts below the agreed threshold during that window, we retrain and redeploy. After the 90-day period you can extend support through a monthly maintenance agreement or run everything in-house.