Arnaud Fouchet

LLM for efficiency: An agentic AI program where we identified several applications that would save employees a lot of time. We worked on concrete cases where the ROI of AI was demonstrated by our POCs, then put our applications into production.

Application examples:

- "Talk to Your Data": A chatbot for making numerical analyses. The key to the adoption of this application was to make the bot's analyses verifiable. To do this, the AI provides queries to get the raw data and reproduce the analysis.

- "Data Reconciliation": Various data sources could name the same product differently. We use LLMs to perform reconciliation. Here, the key was a good visualization of the proposed reconciliations. As AI is not 100% reliable on this subject (in some cases, very understandably), human review was necessary. A simple, pleasant interface was needed, with quick and easy use.

- "Data Cleaning": In a database, addresses were not normalized. To unify everything, we built an agentic AI that used Google Maps to get reliable and structured addresses.

Marketing Mix Model: By cross-referencing sales and marketing data, Pernod Ricard wanted to optimize its advertising campaigns.

In this mission, it was necessary to be pedagogical and clearly explain the model's results.

Then, I participated in the model's redesign, using other mathematical formulations. This was a technically advanced project.

Finally, as a side project, I developed a demand forecasting AI (for inventory management). With the results of my POC, Pernod Ricard decided to launch a whole project with a team of 11 people on the subject.

Managed an R&D team, creating algorithms for online advertising campaign optimization, working on:

- Bidding / pricing: finding the optimal price for ads (machine learning, deep learning)

- Semantics: optimizing ad text (NLP, Word2Vec)

The team managed the entire stack (from algorithm design to server maintenance).

I was also in charge of internal communication (gathering needs from other teams, evangelizing our solutions) and external communication (popularization presentations to clients).

I presented our work at a Python conference (PyData, "How to beat a Google bidder using machine learning")

(Presentation summary: using several methods, including deep learning, transfer learning, and model ensembling, we achieved a bidder that was more profitable for us than Google's. Google's obtained more volume, but ours obtained conversions for less cost and, overall, generated more profit.)

I developed algorithms for SEA campaign optimization (online ads on search engines like Google and Bing):

- Bidding / pricing: algorithms to find the optimal price for our ads (machine learning, data science, python, applied mathematics, statistics)

- Semantics: algorithms to optimize ad text (NLP, Word2Vec)

I handled all project stages (monitoring / state-of-the-art, algorithm development and optimization, production deployment and maintenance).