Integrating generative AI into your product, all the way to production

Useful features, built into the product

AI should fit into a workflow your users already know: producing content, assisting with a task, processing a document. The starting point is the product need, not the model.

• Identify where AI brings a real gain, and where it mostly adds risk.
• Design the integration across backend, APIs and interface.
• Plan for human review when the output requires it.

What keeps an AI feature running in production

• Data: what is sent to the provider, where it is processed, what is retained.
• Reliability: latency, errors, retrying work and behaviour when a provider is unavailable.
• Quality: check results on representative cases before and after each change.
• Usage costs: measure consumption and match it to the service delivered.

Choosing the provider and the architecture

OpenAI, Microsoft Azure and Mistral differ in data hosting guarantees, available models and costs. The right choice depends on your customer and regulatory constraints.

The architecture keeps the option to change model or provider without rewriting the product.

Three references in production

• Dixeo (Edunao): educational content creation and learner support, within learning platforms used at scale.
• ShareAgent: a real-estate agent’s dictation transcribed, then its property features extracted to run a valuation.
• Konta: automatic reading of photographed receipts, confirmed by the user before reaching the books.