Tuka Bade

To be written... stay tuned or call me

To be written... stay tuned or call me

🏢 Companies : CPAM, SFR, Diot Siaci, Midas 👉 Project: Integration of LLMs and AI Agent into Streamlit apps for specifics business use case. 🎯 Context: Within the product engineering teams at SFR and the development teams at CPAM, Diot Siaci and Midas, I trained these teams in using GPT via the API, prompt-engineering techniques and integrating both into FastAPI, Flask or Streamlit applications. ✅ Missions: ➤ Challenge business requirements and define the scope through a reusable technical specification ➤ Integrate prompt-engineering techniques (Chain-of-Thought, zero-shot, one-shot and few-shot learning) ➤ Build a simple chatbot application using FastAPI or Flask ➤ Develop a full chatbot application with prompt and response storage via Streamlit ➤ Define the differences between Machine Learning, Deep Learning and LLMs ➤ Containerise the application using Docker

🏢 Company : ENGIE (1 month) 👉 Project: Creation of a prompt engineering repository to automate content creation. 🎯 Context: For ENGIE’s Legal Research & Development team, I created a knowledge base to enable the business unit to automate content creation, retrieval and verification. I also trained the team to use LLMs and techniques such as zero-shot, one-shot and few-shot learning. ✅ Missions: ➤ Define what AI is and where LLMs sit within the field of AI ➤ Explain how to maintain critical thinking when using an LLM ➤ Advise on how to use LLMs as a legal researcher ➤ Master techniques such as Chain-of-Thought, zero-shot, one-shot and few-shot learning ➤ Demonstrate and embed best technical practices within the team ➤ Explain how to detect biases in outputs and audit a prompt from its formulation to the responses obtained

🏢 Company : ALLEGO (3 months) 👉 Project: Created a data pipeline following a Kappa architecture. 🎯 Context: For Allego, a renewable energy company and European leader in charging networks, I designed a data pipeline using a Kappa architecture to automate the quarterly charging-data submissions that were previously sent manually by the business to a new government service (Qualicharge) via a batch API under the OCPI protocol. The charging dataset comprised both static and dynamic elements: the dynamic data consisted of charging sessions, while the static data detailed charger locations and their charging capacities. ✅ Missions: ➤ Defined and implemented a Kappa architecture on Azure Databricks ➤ Read data from continuously ingested Delta Lake tables ➤ Verified outgoing fields against a reference table ➤ Identified and corrected erroneous fields using validated data ➤ Sent the cleansed data to the government API via HTTP POST requests ➤ Managed HTTP POST sessions with secure authentication ➤ Implemented a logging and alerting system to monitor the data transmissions

👉 Project 1: Classification of visitors to a hotel booking website in order to analyse potential purchasers. 🎯 Context: Traffic data from a hotel group’s website is collected and used to analyse user behaviour and optimise the site journey. For advertising purposes, the Marketing team leverages these data to build targeted ad audiences and promote the hotel’s offerings. Classifying users therefore helps to understand distinct user segments and the common behaviours that drive them to purchase. ✅ Missions: ➤ Audited the existing classification model to determine why it failed to meet marketing requirements ➤ Analysed the dataset to demonstrate overfitting in the current model ➤ Advised the Data team on the scarcity of positive labels (58 million rows, only 0.00418% positive) ➤ Identified correlations between user events to isolate the most relevant features ➤ Built the first draft of the data pipeline, including data cleaning and splitting into training, test and production sets ➤ Retrained the existing Boosted Tree Classifier; trained Logistic Regression and Random Forest models ➤ Calculated feature importances and deployed the Random Forest model into production ➤ Presented findings and delivered workshops to both the Marketing and Data teams

👉 Project 2: Categorisation of users’ purchase probabilities on the website using a clustering model. 🎯 Context: Following on from the user classification project, the Marketing team wishes to categorise the probability segments as follows: very low purchase probability, low purchase probability, moderate purchase probability and high purchase probability. These segments will be used by the Marketing team to create personalised audiences to enhance the performance of their online advertising. ✅ Missions: ➤ Applied arbitrary marketing criteria to define audience segments ➤ Analysed the distribution of model-generated probabilities ➤ Employed a clustering model on the probability scores ➤ Delivered the resulting audience segments into a dedicated table ➤ Presented the audiences and accompanying analyses to the Marketing team