Tadjou T.

Within the Architecture & Innovation team, I design and deploy AI solutions for regulatory compliance and document management. Among its main achievements:

- Implementation of a knowledge base querying system (Confluence, Jira, etc.) via a RAG system, also used to generate commercial requirement documents from regulatory texts.

- Implementation of an MCP Server for a "Chat with Data" use case enabling natural language querying of Redshift and Iceberg databases. I participated in the deployment of these projects on various AWS cloud environments.

Objective: At Gedeon, I worked on setting up a chatbot to answer questions about retirement in France, mainly concerning the Mandatory Retirement Savings Plan (PERO) and related areas.

To achieve this, I performed the following tasks:

- Corpus Construction: Web scraping

- Synthetic Data Creation (FAQ) using OpenAI API prompts and scraped data

- Fine-tuning small LLMs (Llama 7B, Zephyr 7B, Phi 3) on AWS SageMaker with synthetic data and HuggingFace training modules

- Model Evaluation using feedback from retirement experts at Gedeon who annotated the synthetic data and model responses.

- Implementation of the RAG (Retrieval Augmented Generation) approach with three different methods:

- Python code with Langchain, Chroma DB for data indexing and exposing the Chatbot service with Fast API

- Use of Vertex AI Search & Conversation (GCP) with Gemini Pro and DialogFlowCX for a chatbot POC

- Use of Promptflow (Azure) locally with Azure OpenAI and Azure Cognitive connections

- Containerization of the Chatbot designed with PromptFlow for deployment

- Use of GitHub Actions for continuous integration and continuous deployment (CI/CD) of the Chatbot

- Creation of a Service Principal account on Azure Portal for Chatbot deployment

- Chatbot connection via API by implementing a Node.js module in Gedeon's SaaS backend, which communicates with a Mongo DB for saving chatbot conversations

Technical Environment: AWS SageMaker, Notebook, FastAPI, Azure AI Studio, LangChain, LlamaIndex, Python, GitHub actions, Pytorch, HuggingFace, Vertex AI, MongoDB, VSCode

CIFRE thesis on the problem of conversation disentanglement. During this thesis, I worked on issues of building corporate corpora, pseudo-anonymizing data via NER, identifying dialogue acts, and matching text segments in conversations. I also conducted experiments with LLMs, particularly with the LangChain and OpenAI libraries, on some of these tasks. I deployed and used annotation tools such as Prodigy and Inception. After the state-of-the-art on the issues, I developed an Outlook email probe with a graphical interface to optimize email collection while respecting GDPR.