Generative AI and Model Deployment

Organisation over 20 weeks

Four 5-week blocks (4 weeks of classes, 1 week of assessment). The two course units are taken unchanged from the original degree: same content, same assessments.

• Blocks 1 and 2 (weeks 1 to 10): IAG 531 — Generative Artificial Intelligence and Agents
• Blocks 3 and 4 (weeks 11 to 20): MLO 532 — Industrialising and Frugally Deploying Models

IAG 531 - Generative Artificial Intelligence and Agents

To train learners to design applications based on generative models and agents while keeping control of their limitations: factual errors, costs, security risks and technological dependency.

Learning outcomes

• Design an application based on a large language model, choosing between prompt engineering, retrieval-augmented generation and fine-tuning.
• Evaluate the quality, faithfulness to sources and safety of a generative model's outputs.
• Design an agent able to use tools, bounding its autonomy and providing for human oversight.
• Assess the security risks specific to generative applications and propose defensive measures.
• Justify the choice between a locally hosted open model and an external model service in terms of cost, sovereignty and the data processed.

Content

• Generative models: language models, diffusion models, multimodal models; principles, capabilities and limitations.
• Adaptation: prompt engineering, retrieval-augmented generation over a document base, parameter-efficient fine-tuning.
• Agents: tool use, planning, memory, multi-agent orchestration; human checkpoints.
• Evaluation: test sets, human evaluation, faithfulness to sources, hallucinations, biases; evaluation in French and in national languages.
• Security: prompt injection, data leakage, tool misuse; defensive measures, in connection with the web application security risks of the OWASP Top 10:2025 list; tests conducted only in the programme's isolated laboratory.
• Intellectual property in generated content and training data under loi n°2008-09 (Copyright and Related Rights Act); transparency towards users.

Assessment: Continuous assessment (40%): marked practical work. Project (60%): generative application or agent, with an evaluation protocol, a risk analysis and a note on the choice between a local model and an external service. Same arrangements for both cohorts.

MLO 532 - Industrialising and Frugally Deploying Models

To take a model from the notebook to a reliable, monitored and frugal service, able to run where computing power, energy and connectivity are scarce.

Learning outcomes

• Design a model industrialisation pipeline: versioning, testing, deployment, monitoring.
• Detect data or performance drift and organise retraining.
• Optimise a model for constrained deployment through compression, quantisation or distillation, and measure the effect.
• Assess the energy and environmental footprint of an artificial intelligence system over its life cycle and propose a more frugal solution.

Content

• Industrialisation practices: experiment tracking, model registry, continuous integration and deployment, containers.
• Putting into service: inference service, batch processing, monitoring, logging, rollback.
• Production monitoring: data and concept drift, alerts, retraining, user feedback.
• Software quality of a system incorporating a model, according to the quality model of the ISO/IEC 25010:2023 standard.
• Constrained deployment: quantisation, pruning, distillation, lightweight models, edge computing, offline operation and deferred synchronisation.
• Frugality: measuring energy consumption during training and inference, trade-off between performance and cost, hosting choices.

Assessment: Continuous assessment (40%): marked practical work. Project (60%): deployment of a model under constraints, with a monitoring dashboard and an energy assessment comparing before and after optimisation. Same arrangements for both cohorts.

Career opportunities

Artificial intelligence application developer, engineer responsible for model deployment and monitoring.

At your own pace: the self-paced pathway

In a class, this certificate takes 20 weeks, on the cohort's dates. On the self-paced pathway, you start whenever you like, online, and progress at your own pace. The price is the same.

Admission

Bachelor's degree in computer science or a related field, or equivalent experience assessed on application; programming practice and basic knowledge of machine learning. Admission based on application; detailed requirements set by the school.

The credits you earn are recognised without further assessment if you later join the programme they come from, Professional Master's degree in Artificial Intelligence, within a period set by the certificate regulations.