Frequently asked questions about our AI software

By default, all data is stored and processed on Canadian cloud infrastructure (AWS ca-central-1 or Azure Canada Central). If your organisation requires a specific region or on-premises deployment, we accommodate that in the Enterprise tier. Data is encrypted at rest (AES-256) and in transit (TLS 1.2+). We follow ISO 27001-aligned security practices and can provide our security questionnaire responses upon request.

Responsible AI is built into our methodology, not bolted on at the end. For every Growth and Enterprise engagement we conduct a bias audit that examines model outputs across protected attributes such as gender, age and ethnicity. We use established fairness metrics — demographic parity, equalised odds and calibration — and document our findings in a plain-language explainability report. If bias is detected, we apply mitigation techniques such as re-sampling, re-weighting or post-processing adjustments before the model reaches production.

Our core ML stack includes Python, scikit-learn, XGBoost, PyTorch and TensorFlow, depending on the problem type. For data pipelines we rely on Apache Airflow, dbt and Spark. Model serving is handled through FastAPI or TensorFlow Serving behind Kubernetes, with monitoring via Prometheus and Grafana. For NLP tasks we leverage transformer architectures including fine-tuned large language models. We choose the right tool for each job rather than forcing every problem into a single framework.

Yes. Upon project completion and final payment, you receive full ownership of the trained model weights, source code, data pipelines and documentation. We retain no proprietary claim over your deliverables. The only exception is our internal tooling and frameworks, which remain our intellectual property but are licensed to you royalty-free for the duration of your use of the deployed system. This is all spelled out clearly in our client agreement before any work begins.

We set clear, measurable performance targets during the discovery phase — for example, a minimum F1 score, precision threshold or error-rate ceiling. If the model does not meet the agreed benchmarks after reasonable iteration, we work with you to diagnose the root cause, which is most often a data-quality or data-volume issue. In rare cases where the problem proves infeasible with the available data, we provide a detailed technical report explaining our findings and recommendations for a revised approach. We never charge for work that fails to deliver agreed value.