From first conversation to a model running in production, here is exactly what happens and when.
A 45-minute video call, free of charge. You describe the problem, we ask about your data sources, volume and current tooling. By the end we can tell you whether AI is the right approach or whether a simpler solution would serve you better. About one in five enquiries we steer away from machine learning entirely because a rules-based system or a better dashboard would solve the issue faster and cheaper.
We sign an NDA, then you share a representative sample of your data. Our engineers profile it: row counts, missing values, label quality, class imbalance, potential bias. This takes three to five working days. You receive a written Data Readiness Report that scores your data on a five-point scale for each dimension. If the data needs cleaning or enrichment, we outline exactly what is required before modelling can begin.
We build a minimum viable model. For a classification task this might be a gradient-boosted tree trained on your labelled data. For a computer vision task it could be a fine-tuned object-detection model tested on 200 annotated images. The prototype phase is fixed-price: you know the cost up front. It typically lasts two to three weeks. At the end we present accuracy, precision, recall and a confusion matrix, along with sample predictions you can inspect yourself.
Prototypes rarely hit production-grade accuracy on the first pass. We run two or three iteration cycles, each one to two weeks long. Common activities include feature engineering, hyperparameter tuning, data augmentation and, when necessary, collecting more training examples with your team. We share a live dashboard so you can track metrics as they improve. This stage ends when the model meets the agreed performance threshold on a held-out test set.
We package the model as a containerised REST API or integrate it directly into your existing software via a plugin or SDK. Deployment targets we have worked with include AWS, Azure, on-premise Linux servers and edge devices like NVIDIA Jetson boards. We write integration tests, set up logging and configure alerting so you know immediately if inference latency spikes or prediction distributions drift. Deployment usually takes one to two weeks depending on your infrastructure.
Models degrade over time as the real world changes. We monitor prediction quality on a weekly basis and retrain the model on fresh data quarterly, or more often if drift is detected. This ongoing support runs on a monthly retainer. You also get a dedicated Slack channel for questions and a monthly performance summary emailed to your team.
Concrete outputs, not vague commitments.
A 10-to-15-page document covering data volume, quality scores, identified gaps and recommended fixes. Delivered at the end of stage 2.
A reproducible Jupyter notebook with all code, training logs and evaluation metrics. You own this code from day one.
A Grafana or Streamlit dashboard showing model accuracy, latency and data-drift indicators. Updated automatically.
Step-by-step instructions your own engineers can follow to redeploy or roll back the model without our involvement.
A straightforward predictive-analytics project (tabular data, single target variable, fewer than a million rows) moves from discovery call to production in about 47 days. Computer vision projects tend to run longer because image annotation takes time: expect 60 to 80 days. Document-intelligence pipelines sit somewhere in between, usually around 55 days.
These numbers assume your data is already collected and accessible. If we need to help you set up data pipelines first, add two to four weeks to the front of the schedule.
The discovery call is free and takes 45 minutes. We will tell you honestly whether AI is the right tool for your problem.
Book a discovery call