From model to product: my AI Engineer learning path.

Three stages of progress

1. Understand data and compare models

   From pretrained APIs to regression and classification: preparing data, choosing metrics and explaining predictions.

2. Connect application components

   APIs, RAG, storage and MLOps: connecting models to services, tracking versions and experiments, and examining behaviour.

3. Experiment and draw conclusions

   Vision, RL, multimodal data, orchestration and LLM adaptation. Negative results and protocol limitations are part of engineering work.

The projects, one by one

• P1 — Building my AI learning path
• P2 — Clothing segmentation
• P3 — Predicting building energy use
• P4 — Understanding employee attrition
• P5 — Serving a model through an API
• P6 — Tracking credit-scoring experiments
• P7 — Finding events with RAG
• P8 — Deploying and monitoring scoring
• P9 — Scoping an AI project
• P10 — Learning with few labels
• P11 — Training a LunarLander agent
• P12 — Building a multimodal pipeline
• P13 — Orchestrating a chess assistant
• P14 — Adapting a medical LLM

What I want to improve

Strengthen independent evaluations, RAG response faithfulness and served-version consistency. Better document costs, operations and what I can reproduce independently. The P5 authentication competency remains unvalidated in OC feedback.

My full-stack experience helps me consider the whole application. Training pushes me to make data, model and protocol choices explicit, including their limitations.

How my view of the profession changed

Before training, I mainly saw the job as integrating AI into applications. I discovered the work needed to prepare data and understand the business problem before choosing a solution. Today I consider the entire chain: needs, data, models, evaluation, deployment and monitoring.

Technical choices must serve product performance, costs and security. This requires a broad perspective, ongoing learning and the ability to adjust multiple components. Small open-source models tailored to specific needs are another direction I want to explore.

Two projects that helped me progress

Building consumption: understand before comparing

I spent considerable time understanding models, metrics and what distinguishes sound evaluation from misleading results. This project developed my feature engineering and data leakage prevention skills. If I repeated it, I would handle missing values and outliers more carefully, focus more on feature selection and creation, and compare several hyperparameter tuning approaches.

Scoring and MLOps: connect a model to a service

Deploying the scoring model connected modelling work with an API, unit tests, versioning, containers and CI/CD. I better understood why deployment and monitoring matter as much as training. Reinforcement learning also interested me, particularly analysing agent behaviour and testing adjustments.

My priority: stronger RAG and agent evaluation

I want to improve retrieved-document relevance and answer faithfulness to sources. My goal is to build tests that compare versions and establish whether a change genuinely improves answers. This is a development priority, not a capability already demonstrated by past results.

Explain choices and understand my strengths

In my daily work, I explain AI uses at different levels of detail. I first justify a solution with an accessible comparison, then go deeper if requested. For example, explaining why an LLM benefits from sources, tools and checks makes RAG design choices easier to understand. These components help reduce and detect errors; they do not guarantee a correct answer.

I identify with clear communication, curiosity, independence and patience. Building this portfolio helps me identify my skills and limitations more clearly, and better explain my projects and technical decisions.

Method and evidence

The detailed report uses OC feedback and inspected public repository files. A recorded result, assessor feedback and current functioning are different evidence. No project was run during this audit.

Fourteen validated projects. P14 status was updated following Pierre’s confirmation on 11 October 2026; technical evidence remains from the 9 October audit. Pierre confirmed the award of the RNCP level 7 certification on 11 October 2026.