YOUR ROLE
As part of your studies, you are looking for a 4- to 6-month internship to complete your degree (Master's 2, engineering school or equivalent).
You will join a Data Science team working on an applied research project: RAFA, an artificial intelligence agent that automates part of the data scientist's experimental cycle. RAFA designs, codes, runs, repairs and analyses machine learning experiments, with a particular focus on multi-horizon time series forecasting applied to shipping lines.
Your role will be to make RAFA a more reliable, robust and useful tool: improving the agent's experimental logic, validating results, preventing data leakage and integrating advanced Agentic AI and Deep Learning models.
WHAT WILL YOU DO?
- Experimentation logic : You will improve how model families are selected and structured, the parent/child logic of the experiment tree and the reuse of previous results, so that RAFA progressively learns which experiments are relevant and can fall back on a robust baseline.
- Forecasting & backtesting : You will work on multi-horizon forecasting (time-based supervised datasets, lags, lagged features, per-horizon metrics) and rigorous backtesting, with one key goal: beating a strong baseline through residual models, baseline/model gating or cautious baseline correction.
- Agentic AI & multi-agent orchestration : You will help orchestrate RAFA's specialised agents (analyst, planner, coder, repairer, reviewer, validator, reporter): robust prompts, experiment memory, automated validation, anti-leakage guardrails and a reliable execution graph.
- Applied Deep Learning : You will integrate and evaluate neural models (MLP with embeddings, TCN, RNN/LSTM/GRU, temporal Transformers, hybrid baseline + correction models) to identify where deep learning brings real gains over classical models.
- Code generation & repair : You will make the conversion of an experiment plan into executable Python code more reliable: better coder prompts, static checks, detection of temporal leakage, analysis of failure causes and fewer unnecessary repair loops.
- Reporting & documentation : You will produce HTML reports, experiment-tree visualisations, technical documentation of the agent and summaries for data scientists and management.
WHO ARE WE LOOKING FOR?
- Education : You are a Master's 2 or engineering school student (or equivalent), specialising in data science, machine learning, artificial intelligence, statistics or time series.
- Technical skills : You are proficient in Python (pandas, NumPy, scikit-learn), Jupyter and Git. You have solid foundations in supervised machine learning, forecasting, time-based backtesting and regression metrics (WAPE, MAPE, MAE, RMSE), and a genuine awareness of data leakage issues.
- Agentic AI : You have initial experience with LLM agents and orchestration frameworks (LangGraph, LangChain or equivalent), prompt engineering and automated code generation.
- Deep Learning : You are comfortable with PyTorch and time series architectures (MLP, TCN, RNN/LSTM/GRU, Transformers), as well as training techniques (regularisation, early stopping, loss optimisation).
- Nice to have : Dataiku, FastAPI, Docker, MLflow, OpenAI API or equivalent, NetworkX/Graphviz, LLM observability.
- Languages : Professional English required.
- Personal qualities : You are curious, rigorous and autonomous, with a strong interest in AI agents and exploratory topics. You can analyse experimental results and document your work clearly.
JOB ENVIRONMENT
- Location : MARSEILLE
- Internship : 6 months from January
Come along on CMA CGM's adventure !