Ludovic Kun

As a Senior Data Scientist and Consultant, I interacted with clients to understand their business needs and provide them with tailored A.I solutions: ⚡ Built and deployed advanced NLP pipelines (NER, document classification, topic modeling, trend detection) using cutting-edge transformer models (RoBERTa, DeBERTa, LLaMA). ⚡ Developed intelligent chatbots and agentic applications with GPT-4, Claude, and RAG for legal, customer support, and product recommendation use cases. ⚡ Fine-tuned large language models (SFT, LoRA) for domain-specific tasks, improving accuracy and adaptability. ⚡ Delivered AI solutions across industries (logistics, healthcare, services, retail), streamlining workflows and saving clients several workdays per month.

As a Deep Learning engineer, I designed and deployed machine learning solutions to support the detection of misinformation and security threats at scale: ⚡ Built multimodal data processing pipelines to extract signals from text, images, video, and audio, including OCR and multilingual speech-to-text transcription, leveraging CNNs and Transformers (PyTorch). Processed datasets exceeding millions of documents and media items. ⚡Collaborated on the development of a large-scale misinformation detection system for TikTok, integrating linguistic and visual features. The solution scaled to analyze thousands of videos daily, reducing manual moderation workload by ~40%. ⚡Developed end-to-end ML solutions covering dataset creation, model training, evaluation, benchmarking, and experiment tracking (MLFlow). Achieved up to 30% improvement on key metrics (F1-score, WER, etc.) over baselines. ⚡Implemented specialized computer vision models for object detection (weapons), landmark recognition for geolocation, and emotion recognition (EfficientNet) using PyTorch and TensorFlow. Achieved near real-time inference (<200ms latency). ⚡Prototyped and deployed models via REST APIs (Flask, FastAPI) and interactive demo interfaces (Gradio), enabling seamless testing and adoption by non-technical stakeholders. Delivered measurable operational impact: deployed AI solutions that assisted fact-checkers in analyzing cross-platform content daily, doubling throughput of verified items and cutting response times during critical events (e.g., elections, crisis monitoring).

As a A.I engineer, I designed end to end solutions to detect risks on social media platforms: ⚡Developed and deployed Deep Learning solutions using PyTorch, which detect various types of risks on social media (texts, images, videos). ⚡Pretrained, fine-tuned, and evaluated models (CNN, BERT, Electra, RoBERTa) for text and video data (sentiment analysis, video classification, risk classification). Processed datasets exceeding 10M+ documents and media items. ⚡Pushed the limits of deep learning models using strategies such as knowledge distillation, ensembling, and self-ensembling, enabling deployment of smaller models with strong performance. ⚡Served models with TorchServe and deployed them on GCP (Docker, Google AI Platform). Models have been deployed, allowing moderators to filter hundreds of harmful content items per day.

Enhanced the user experience for Windows customers by analyzing Input Session Telemetry data and developing rule-based detection algorithms. ⚡Collected, preprocessed, and analyzed large-scale telemetry datasets using Scope/SQL, designing efficient queries for data exploration and anomaly detection. ⚡Identified unusual patterns and anomalies in keyboard input sequences through statistical analysis and rule-based methods, leveraging Python (pandas) and Jupyter Notebooks. ⚡Collaborated with engineers to investigate anomalies in Visual Studio, presenting findings and actionable insights to the team for product improvement. Delivered measurable impact: detected critical telemetry issues on Japanese keyboards, leading to fixes that improved the Windows user experience for thousands of users.

Designed and delivered a Proof of Concept (PoC) for detecting and recognizing cosmetic defects on optical lenses using Deep Learning. ⚡Researched and evaluated state-of-the-art object detection algorithms (e.g., Faster R-CNN, YOLO, SSD), selecting the most effective approaches for industrial defect detection. ⚡Implemented and benchmarked models in TensorFlow/Keras, achieving higher accuracy in defect classification and localization compared to previous baseline. ⚡Developed a demonstration GUI in C#, enabling real-time interaction with the model in a production-like environment and facilitating stakeholder validation. PoC success led to investment in a production-ready system now filtering thousands of lenses daily, significantly improving quality control throughput and consistency. A patent application based on this work is currently under review.

Developed a computer vision solution to estimate age and gender from facial images. ⚡Researched state-of-the-art approaches in facial analysis and image classification, experimenting with CNN-based architectures (VGG, ResNet, Inception) to identify morphological criteria. ⚡Created and labeled training and test datasets, applying strategies to mitigate dataset bias and improve generalization across demographics. ⚡Implemented and benchmarked multiple deep learning models using the MXNet framework in Python, achieving robust performance on age and gender prediction tasks. ⚡Delivered production-ready outcomes: the solution was integrated into the client’s application (Facelytics) and deployed in retail analytics systems to analyze customer demographics in shopping malls, supporting data-driven marketing strategies.

Proposal : Modeling and simulation of humanoid robots for improved balance in the presence of external force disturbances. Simulation of a humanoid robot (HOAP-2) in order to study its stability. Implementation and study of algorithms about the stability of robot. Using matlab, OpenHRP3 and C ++.

- Migrating a database - Analysis and correction of customer data (using mainly Excel)

Implementation of synthesis models with graphics that compare different years depending on different variables (turnover, gross values ...)