GenAI at scale: from experimentation to automation

The context

Several GenAI initiatives had emerged around varied business problems: document understanding, classification, call analysis, information extraction, decision support and conversational interfaces.

The challenge was to turn these experiments into solutions that could genuinely be operated in production: reliable, integrated with existing processes and able to evolve with business needs.

I contributed technically to the industrialisation of more than 10 GenAI use cases, working across application development, production deployment and ongoing evolution.

From experimentation to automation

The use cases covered several families of problems:

• document understanding and classification;
• structured information extraction;
• call transcription and analysis;
• automation of control and verification tasks;
• decision support in business processes;
• conversational interfaces;
• autonomous agents orchestrating business steps.

The goal was not simply to integrate a language model, but to build solutions that could fit sustainably into existing processes.

A cross-functional GenAI capability

The projects combined different approaches depending on the need: language models, retrieval-augmented generation, document understanding, speech processing and specialised agents.

This diversity requires an engineering approach rather than a simple experimentation mindset.

Each solution has to meet several constraints simultaneously: result quality, robustness, application integration, performance, observability and ability to evolve.

My role

I contributed to the development and hardening of the GenAI solutions, with particular involvement in getting them into production.

My work included:

• application development and evolution;
• fixing critical bugs and regressions;
• architecture evolution and refactoring;
• hardening deployment pipelines;
• data model design and evolution;
• schema migrations and evolution;
• optimisation of execution environments;
• analysis and resolution of production incidents.

This experience led me to work across the full chain, from experimentation to operating GenAI solutions in production.

Concrete results

More than 10 GenAI use cases were industrialised and deployed to production, covering document automation, classification, call analysis, structured extraction and autonomous agents.

For some processes, AI significantly increased the level of automation and reduced manual intervention.

Another document-extraction use case achieved high rates of usable extraction on highly structured documents, with output directly consumable by business processes.

Moving to production changes the nature of the problem

A GenAI experiment can work with a handful of examples and a limited objective.

A production solution must withstand changing data, new business cases and model evolution.

A prompt change can improve one case and degrade another. A model upgrade can change application behaviour. A performance optimisation can introduce a new functional constraint.

That is why AI Engineering quickly meets the same concerns as Software Engineering: versioning, testing, deployment, observability, regression management and operations.

What I take from it

Industrialising GenAI is not simply about choosing the right model.

Value appears when models are integrated into systems that can operate sustainably: applications, data, processes, monitoring and teams.

It is this convergence of AI Engineering, Software Engineering and business understanding that turns a promising experiment into an industrial capability.