FAQ
Answers to your
ML questions
Everything worth knowing about ML development, data, pricing, and support.
General
What does Codeworth do?
▼
Codeworth is an ML/AI consultancy. We build custom machine learning models, fraud detection systems, NLP pipelines, RAG chatbots, and computer vision solutions — then deploy them to production with full MLOps infrastructure.
Which industries do you work with?
▼
Our main sectors are FinTech (fraud detection, credit scoring), Healthcare (clinical NLP, triage automation), Retail (demand forecasting, recommendation engines), Manufacturing (quality control CV), and SaaS (churn prediction, LLM features).
Where are you based? Do you work remotely?
▼
We are a UK-based ML consultancy serving clients in the UK, EU, and globally. All collaboration happens remotely via Zoom, Slack, Notion, and GitHub.
Can I see examples of your ML work?
▼
Yes — our portfolio has 40+ ML case studies. Each case shows the business problem, data used, ML approach, metrics achieved (F1, precision, recall, ROI), and the MLOps setup.
ML Process
What does a typical ML project look like?
▼
Every project follows five phases: (1) Discovery — understand the business problem and data availability; (2) Data Audit — assess quality, labelling needs, and gaps; (3) Model Training — iterative experiments with MLflow tracking; (4) Production Deploy — FastAPI serving, Docker/K8s packaging, CI/CD; (5) MLOps — drift monitoring, automatic retraining, documentation handover.
How long does an ML project take?
▼
A proof of concept runs 3–6 weeks. A full production deployment with monitoring takes 8–16 weeks. The primary variable is data readiness — projects where you already have labelled data move significantly faster.
What metrics do you guarantee?
▼
We specify minimum acceptance metrics in the project spec before starting (e.g. precision ≥ 0.90, recall ≥ 0.85, AUC ≥ 0.92). The model is considered production-ready only when it meets those thresholds. If it doesn't — we keep training at no extra charge.
Do you do the data labelling too?
▼
Yes — we can manage the labelling process using Label Studio, Prodigy, or third-party annotators. Labelling costs are estimated separately based on dataset size and annotation complexity.
What happens after the model is deployed?
▼
We hand over all code, model artefacts, API specs, MLflow configuration, and documentation. We also offer ongoing MLOps retainer packages for drift monitoring, automatic retraining, and version management.
Data & Privacy
Do you need access to our production data?
▼
Not necessarily. We can work with anonymised or synthetic data for development and use real data only for final validation. For sensitive sectors (healthcare, finance) we support on-prem compute so data never leaves your infrastructure.
How do you handle GDPR compliance?
▼
We sign a Data Processing Agreement (DPA) before any data exchange. Training data is stored only as long as needed. On-prem deployments are available for maximum data sovereignty. We document all data flows for your compliance records.
Can the model run on our own servers (on-prem)?
▼
Yes. We support on-prem, private cloud (AWS/GCP/Azure VPC), and hybrid deployments. For regulated sectors this is often the preferred option — your data and model weights never leave your environment.
Who owns the model and code after delivery?
▼
You do. After full payment all code, model artefacts, training notebooks, and MLflow experiment logs transfer to you. We retain the right to mention the project type in our portfolio (without client data).
Pricing
How much does ML development cost?
▼
A proof of concept starts at £1,800 (3–4 weeks). A full production model with API and MLOps monitoring starts at £4,500 (8–12 weeks). Enterprise multi-model systems are priced individually after a discovery call.
What payment terms do you offer?
▼
50% upfront after signing the contract, 50% on final delivery. For larger engagements we can split into three milestones: 30% / 40% / 30% aligned to discovery, prototype approval, and production deploy.
Is there a cheaper way to start?
▼
Yes — our Proof of Concept package (£1,800) is designed exactly for this. You get a working prototype with documented performance metrics. Most clients use it to validate the ML approach before committing to a full production build.
Do you offer ongoing maintenance pricing?
▼
Yes — MLOps retainer packages start at £800/month. They include drift monitoring, automatic retraining triggers, monthly performance reports, and up to 4 hours of model updates.
Compliance & Regulation
What is FCA SS1/23 and why does it matter for ML?
▼
FCA SS1/23 is the Financial Conduct Authority's Supervisory Statement on Model Risk Management published in 2023. It sets out expectations for UK financial firms on model governance, validation, documentation, and ongoing monitoring. Any ML model used for credit decisions, insurance pricing, or trading at a regulated firm must comply. Codeworth delivers model documentation packages meeting SS1/23 requirements.
Do we need a DPIA for our ML project?
▼
A Data Protection Impact Assessment is mandatory under UK GDPR for high-risk ML processing: large-scale profiling, systematic monitoring, or processing sensitive categories (health, financial, biometric data). Most production ML models in fintech, healthcare, and HR require a DPIA. We complete DPIAs as part of our project delivery — you receive a ready-to-submit document for your DPO.
How does UK GDPR Article 22 affect ML systems?
▼
UK GDPR Article 22 gives individuals the right not to be subject to solely automated decisions with significant effects (loan refusal, insurance denial, hiring rejection). You must provide: a human review option, meaningful explanation of the decision, and ability to contest it. Codeworth builds explanation modules and human review workflows into every regulated ML system.
What is Consumer Duty and how does it affect ML?
▼
FCA Consumer Duty (July 2023) requires financial firms to demonstrate good customer outcomes. ML models used for pricing, product selection, or customer communications must be shown not to harm retail customers — including vulnerable customers. Codeworth conducts Consumer Duty impact assessments for ML systems and documents outcomes monitoring frameworks.
How do you ensure ML models are not biased against protected characteristics?
▼
We run disaggregated performance analysis across UK Equality Act protected characteristics (or proxies), apply SHAP fairness analysis to detect proxy discrimination, and document our bias testing methodology in the model card. For FCA-regulated models, we follow the FCA's fairness in financial services ML guidance.
Technology & Integration
What cloud platforms do you work with?
▼
We are cloud-agnostic: AWS (primary — SageMaker, Lambda, RDS, S3), Microsoft Azure (Azure ML, Cognitive Services, Cosmos DB), Google Cloud (Vertex AI, BigQuery, Cloud Run). For UK regulated industries requiring data residency, we deploy to AWS eu-west-2 (London) or Azure UK South. We also support on-premises deployment for NHS and defence clients.
Can your ML integrate with our existing software?
▼
Yes. We integrate with CRM (Salesforce, HubSpot, Dynamics 365), ERP (SAP S/4HANA, Oracle ERP Cloud, Sage), data warehouses (Snowflake, BigQuery, Redshift, Azure Synapse), and business applications via REST APIs, webhooks, or native connectors. We document all integration points in technical architecture specifications.
What is your standard ML tech stack?
▼
Python 3.12, scikit-learn, PyTorch or TensorFlow, FastAPI for model serving, MLflow for experiment tracking, Airflow for pipeline orchestration, Docker and Kubernetes for deployment, Great Expectations for data quality. We adapt the stack to your existing infrastructure — if you're invested in Azure, we use Azure ML and Databricks rather than AWS.
How long does it take to deploy a production ML model?
▼
PoC: 3-6 weeks. Production build: 8-16 weeks. Full MLOps pipeline including monitoring and retraining: add 2-4 weeks. Timeline depends on data readiness, integration complexity, and regulatory documentation requirements. NHS and FCA-regulated projects add 2-4 weeks for compliance documentation and review.
Do you provide model documentation and handover?
▼
Yes. Every production ML delivery includes: model card (intended use, performance metrics, limitations, training data description), technical architecture diagram, API documentation, MLOps runbook, and FCA-compliant model risk register entry. We also provide 4 weeks of hypercare support after go-live and optional ongoing MLOps retainer from £800/month.
MLOps and Infrastructure
What is MLOps and why does it matter?
▼
MLOps is the set of practices that deploys, monitors, and maintains ML models reliably in production. Without it, models degrade silently as data distributions shift. For UK regulated industries, MLOps is not optional — FCA SS1/23 and NHS audit trail requirements both demand documented, reproducible model lifecycle management. Codeworth builds MLOps tooling into every production ML project as standard.
Which cloud platform should we choose for UK ML workloads — Azure, AWS, or GCP?
▼
The right choice depends on your regulatory context and existing tooling. Azure UK South is the preferred option for FCA-regulated workloads because it satisfies UK data residency requirements and integrates natively with Microsoft 365 compliance tooling. AWS offers the broadest ML service catalogue (SageMaker, Bedrock, Rekognition) and is strongest for general ML engineering. GCP and Vertex AI suit research-oriented teams and BigQuery-heavy data stacks. NHS Digital projects typically require Azure UK South or AWS eu-west-2 (London) for data residency. Codeworth is cloud-agnostic and recommends based on your specific compliance, cost, and integration requirements.
How much data storage does an ML project typically require?
▼
Storage requirements vary significantly by model type. NLP models with text datasets typically need 10–50 GB. Computer vision projects with image or video data commonly require 100–500 GB or more. Tabular models for fraud detection or credit scoring are usually under 10 GB. The critical factor is not volume but data quality — a clean, well-labelled 5 GB dataset frequently outperforms a noisy 100 GB one. We conduct a two-week data readiness audit at project start to assess what you have, what is missing, and what needs cleaning before modelling begins.
What is model monitoring and how frequently should it run?
▼
Model monitoring tracks data drift (changes in input feature distributions) and model performance degradation (drop in precision, recall, or AUC on live data) after deployment. Monitoring frequency depends on your risk profile: FCA-regulated models (credit, insurance) require daily automated checks with alert thresholds defined in the model risk register. NHS clinical models typically run on weekly monitoring cycles. Standard commercial models are monitored monthly. All Codeworth MLOps setups include automated alerting via Slack or email when drift or performance metrics cross agreed thresholds.
Can you migrate our existing on-premise ML models to the cloud?
▼
Yes. We have experience migrating on-premise model serving environments to Azure UK South, AWS eu-west-2 (London), and hybrid configurations where sensitive data remains on-prem while compute scales in the cloud. A typical migration takes 4–8 weeks and includes model re-validation in the new environment, infrastructure-as-code documentation, and a data residency confirmation sign-off. All migrated models undergo performance regression testing to confirm cloud results match on-premise baselines before go-live.
Data and Privacy
How does Codeworth handle our data under UK GDPR?
▼
We act as a data processor under UK GDPR. Before any data transfer we sign a Data Processing Agreement (DPA) that specifies lawful basis, retention periods, and sub-processor chains. Your data is used exclusively for the contracted project and is not used for any other client engagement or to train internal models. All data is deleted or returned at project completion in accordance with the DPA. We maintain a UK GDPR-compliant data register and can provide documentation for your own compliance records or DPO review.
What is the minimum amount of data needed to build an ML model?
▼
Minimum viable data depends on the task. Binary classification models typically need at least 1,000 labelled examples per class, though 5,000+ per class is preferable for production accuracy. Time series forecasting models require at least two full years of historical data to capture seasonal patterns. NLP text classifiers can work with as few as 500 labelled examples per category when combined with active learning and pre-trained transformer fine-tuning. We complete a data readiness assessment at the start of every engagement so you know exactly what you have and what gaps need addressing before modelling begins.
Can you work with synthetic or anonymised data?
▼
Yes. For sectors where real data cannot leave the organisation we use synthetic data generation techniques including CTGAN (Conditional Tabular GAN) and differential privacy methods to produce statistically representative training sets. Anonymised financial datasets approved for NHS and FCA-regulated use cases are also available for benchmarking and pre-training. We document the synthetic data generation methodology in the model card so your compliance team has full audit trail visibility.
Do you sign NDAs, and how is intellectual property handled?
▼
Yes. We sign mutual NDAs at project start before any confidential information is exchanged. All intellectual property arrangements are specified in the project contract: the client owns 100% of all code, trained model weights, training datasets, and documentation produced during the engagement. Codeworth retains no ownership or licence rights. We may reference the project type and sector in our portfolio as an anonymised case study unless you request full confidentiality, in which case we sign a non-reference clause.
What happens to our data and models if Codeworth stops operating?
▼
Our standard DPA includes data return and destruction provisions that apply regardless of the reason for contract termination, including business cessation. All code is held in client-owned or client-accessible repositories throughout the project, not just at handover. Trained model weights and artefacts are stored on client infrastructure or transferred to client-controlled cloud storage. In the event of business discontinuation, we commit to a 30-day wind-down period to complete data transfer and model handover. These provisions are enforceable contract terms, not informal commitments.
Support
What warranty do you provide?
▼
All projects include 3 months of warranty support after production deployment. If the model performance drops below the agreed thresholds for reasons within our scope — we fix it at no extra cost.
How does ongoing support work?
▼
Via Slack or email — we respond within 4 business hours. For critical production issues (model down, serving errors) we target a 2-hour response and same-day resolution.
Can your team train our in-house data scientists?
▼
Yes. We offer knowledge transfer sessions covering model architecture decisions, MLflow setup, deployment configuration, and monitoring dashboards. We also provide annotated notebooks and written runbooks.
What if our data distribution changes significantly after deployment?
▼
This is called concept drift — it's expected and normal. Our MLOps setup monitors data and model performance metrics continuously and alerts you when drift exceeds a threshold. The MLOps retainer includes automatic retraining triggers for this scenario.
Ready to launch
your project?
Leave a request and get a free ML consultation within 24 hours. Together we'll identify which solution will deliver the highest ROI.