We developed this process over roughly 40 client projects. The phases overlap slightly in practice, but the gates between them are firm: we do not move forward until both sides agree the current phase is complete.
1
Discovery call and data audit
This is a 60 to 90-minute video call. We ask about the specific task you want to improve, who does it today, how long it takes, and what data you collect along the way. We also ask about your tech stack: where your data lives, what format it is in, and who has access.
After the call we review a sample of your data (under NDA) and write a one-page feasibility note. That note says one of three things: yes, this is a good fit for AI; yes, but you need to fix a data-quality issue first; or no, a simpler solution would serve you better. We have told roughly one in five prospective clients that they did not need us. We would rather be honest early than bill for work that will not deliver.
2
Scope and proposal
If the feasibility note is positive, we write a detailed proposal. It specifies the model type, the input data, the target metric (for example, mean absolute percentage error for a forecast, or F1 score for a classifier), and the threshold we consider success. It also includes a fixed price for the build phase and an estimated timeline, typically four to eight weeks.
We do not pad proposals with optional extras. The document covers one problem, one deliverable, one price. If you want to tackle a second problem later, that becomes a separate project.
3
Data preparation and feature engineering
This phase usually takes the longest. Raw business data is rarely clean enough to train on directly. We write extraction scripts, handle missing values, normalise units, and create derived features that help the model learn patterns. For a demand-forecasting project, that might mean generating day-of-week flags, rolling averages, and holiday indicators from your order history.
Your team stays involved here. We schedule a weekly 30-minute sync to confirm that our assumptions about the data match reality. A column labelled "status" might mean something different to your warehouse team than it does to your finance team. These conversations save weeks of rework.
4
Model training and validation
We train multiple model architectures on your prepared data and compare their performance against the metric defined in the proposal. For tabular data we typically test gradient-boosted trees, linear models, and sometimes a small neural network. For text tasks we fine-tune transformer models on your labelled examples.
Validation uses a holdout set drawn from a time period your model has never seen. We do not report accuracy on training data because that number is meaningless in production. The validation report includes error distributions, edge-case analysis, and a plain-English summary of where the model struggles. If performance falls below the agreed threshold, we either iterate on features or recommend stopping the project. You do not pay for a model that does not meet the bar.
5
Deployment and integration
A model sitting in a notebook is not useful. We package the trained model as a REST API or embed it directly into your existing workflow tool. Common integration targets include Google Sheets (via Apps Script), Salesforce, internal dashboards built on Retool or Metabase, and ERP systems like SAP or NetSuite.
We handle authentication, rate limiting, logging, and error alerts. Your team gets a runbook that covers what to do if the API returns an error, how to check model health, and who to call if something looks wrong. Deployment typically takes one to two weeks depending on the complexity of your environment.
6
Monitoring and retraining
Models degrade over time. Customer behaviour shifts, product lines change, and seasonal patterns evolve. We set up automated drift detection that compares incoming predictions against actual outcomes on a rolling basis. When accuracy drops below a configurable threshold, the system flags it and we schedule a retraining cycle.
Retraining is covered under an optional monthly support plan. The plan also includes priority access for bug fixes, minor feature changes (like adding a new input column), and a quarterly performance review call. Most clients stay on the support plan for at least 12 months. Some projects have been running for over two years with only minor adjustments.
