Elias L.

Second mission, directly under the supervision of the Director of Data Science:

- Developed several PoCs, targeting production deployment after Executive and Director Committee validation

- Worked in a 3-person squad (1 PM, 2 Data Scientists), with fast and efficient approach that fitted perfectly the needs to demonstrate our capabilities to create useful agents

- Built SET 2.0, a PoC to automate the first-level analysis of retail opportunities for Carrefour (creation, extension, or acquisition of stores)

- Used a multi-agent approach to evaluate key factors: competition, cannibalization risk, and topography

- Used several technologies such as Agent Development Kit, Cloud Run, LangChain, geo-data calculus (for catchment areas and topography studies)…

- Weekly meetings with each stakeholders to increase adoption and make sure to fit the business needs

First mission, with the Pricing team:

- Worked on the chaining algorithm of products to determine their optimal price

- Conducting experiments, workshops and deployments to maintain and improve the current solution, while exploring new ones

- Leading a GenAI experimentation with a Google Blackbelt to create a « Pricer » Gemini agent, then switching to a hybride architecture (classic ML algorithm + LLM for ambiguous cases)

- GCP stack: Vertex AI (Datasets, AutoML…), Big Query (Data Form included), Composer

- Relations Extraction with Few-Shot Learning Model, by exploring and optimizing state-of-art BERT vector combination techniques, reaching a minimum of 75% accuracy on complex real world data.

- Making a NER Few-Shot model for a PoC, with deep debugging and analysis, reaching an overall ≈80% F1-score on OCRised production data (very noisy).

- Deep Learning model in Java for Sentence Segmentation, to improve a custom rule-based NER solution.

- Dataset Generation with auto-labelling using Generative AI (LLMs), to reduce the time-to-market by quickly fine tuning production models to customers-like documents

- Designing and creating a whole pipeline about Active Data Generation (automating the Active Learning technique), to finetune a NER and a RE model automatically using a sequential augmentation technique with GenAI and a multi-agent approach until reaching threshold results

Assistance of a researcher to study the capabilities of Distance Metric Learning for unsupervised Out-of-Distribution Detection (OOD).

- Defining and developing the research questions, and the goals of the project.

- Conducting a State-of-Art about the topic.

- Implementing solutions and running experiments.

- Writing reports in Latex for a latter integration in a publication, and Power Points presentations.

- Attending to research events, such as ICT.OPEN 2023 or the CWI Machine Learning Theory Boot Camp.