Adrien Biarnes

- Led the design and training of HSTU-based generative foundation models over merchant event sequences, scaling to billions of tokens and tens of millions of merchants, and turning what was historically a fleet of bespoke supervised models into a single shared sequence backbone. - Drove the GPU performance work end-to-end: authored and tuned custom Triton attention kernels (variable-length, jagged-sequence, and ragged-causal variants) to make HSTU-style attention tractable on long merchant histories, materially improving training throughput and unlocking longer context windows than off-the-shelf FlashAttention paths allowed. - Prototyped and benchmarked hybrid architectures beyond pure transformers — including Liquid-AI–style state-space/liquid time-constant blocks interleaved with attention — to evaluate compute/quality trade-offs on long, bursty merchant timelines where pure self-attention is wasteful. - Translated the foundation model into concrete merchant-facing products: zero/low-shot automated merchant selection for paid marketing campaigns, plan-upgrade and plan-downgrade propensity predictions, churn and lifecycle signals — all served from shared embeddings rather than re-trained per task. - Built the causal evaluation and debiasing layer on top of the model: combined IPW / doubly-robust estimators with the foundation-model embeddings as nuisance features to produce counterfactual estimates of treatment effects (campaign eligibility, plan offers) that correct for the heavy selection bias in historical merchant outreach data. - Owned the research → production loop: training infra on multi-node H100s, evaluation harnesses for both predictive and causal metrics, and the rollout path into Shopify's marketing and growth surfaces.

* Design and implementation of a new user experience for the news website and mobile applications (https://ici.radio-canada.ca/info/mon-fil). - New personalized recommendation feed for connected users. - Handled the project from start to finish, from requirements gathering to deployment and monitoring in production, including implementation of the multi-stage recommendation pipeline (candidate generation/ranking). * Maintenance and evolution of an anonymous session-based recommendation system (based on the gru4rec architecture) embedded into various applications of the company. Tools: Python, Docker, Catboost, GCP (BigTable, Cloud Functions, Kubeflow, Vertex AI, MemoryStore)

- Research and development focused on recommendation algorithms at Dailymotion: * Collaborative model: a deep neural network based on a recurrent architecture (Gru4Rec based). * Content-based model: search for nearest neighbors in a semantic embedding space (NLP) * Multi-stage recommendation strategy (candidate generation/filtering/ranking) - Iterative approach on the following main axes: * Signal enhancement (data mining) * Feature engineering * Model evaluation (off-policy learning/evaluation) * Explicability of models - Numerous production releases with positive AB tests => multiplied by 2 fold the main engagement metrics during the course of my tenure. Tools: Python, Tensorflow, Airflow, MLFlow, DataFlow, BigQuery, Docker, Kubernetes, Scikit-learn, GCP

Design and implementation of a pipeline to perform AB testing of bidding algorithms on adwords

- Development of computer vision, machine learning and deep learning algorithms applied to television advertising (classification of video sequences, video search engine, logos detection, text extraction) - Maintenance and evolution of a system for commercials audio identification using signal processing technics (spectral signature based) on a wide range of TV channels (mainly french) - Setting up processes for the evaluation, the training (on a regular basis) and deployment of models in production (Azure and AWS) - Technical lead, recruitment and management of employees within the data team - Presentation of a part of my work at a local meetup of video processing experts Technical environnement: Python, Scitkit, PyTorch, Tensorflow, Celery, Flask, AWS, Azure, RabbitMQ, OpenCV, FFMPEG, SQL, Git

- Estimation of price elasticity of demand for a major telecommunication operator. Comparison of econometric estimators adapted to panel data (fixed and random effects, first difference, generalized linear models) versus machine learning algorithms. Time series clustering. - Factor analysis for a transportation company and an actor in the energy sector. Technical environnement: R, Shiny, Python, Keras

Development in a team of 6 to 7 people in AGILE mode (short cycle / frequent deployments) on a service oriented distributed architecture. Evolutions or complete overhaul of the different modules for the internal teams in the following areas: - Risk management - Thirdparties financial reconciliation and accounting - Internal data referential design (corporate actions, closing prices, security lifecycles) - Trading fees configuration Development of transversal modules: - Administration and configuration of internal applications - Enterprise service bus middleware - Monitoring of the overall architecture and continuous deployment Technical Environment: C#, Sql Server, WPF, Winforms, Redis, ZeroMQ, Cassandra, Visual Studio, Resharper, Mercurial

* Development of the eFront transversal platform: - Javascript front-end framework (portal construction, client-server communication, UI design modules) - Platform configuration modules - Analytical platform (cubes servers, dashboards) - Synchronisation module with Microsoft Exchange servers * Consulting missions for different clients: - Code and architecture audit - Platform expertise for general debugging * Development of a customer relation management system for AXA France and La Médicale Technical environment: .NET (C# and VB), Javascript, Visual Studio, Sql Server, Oracle, Exchange

Consultant for the french army : - Functional and technical specifications - Developing in Java/J2EE - Project management - Unit tests and validation tests Technical environment: Java JEE (JSF, Spring, Hibernate), JBoss Portal et AS, SQL (Oracle), Eclipse, Maven 2, JUnit

Design and Development of a proof of concept based on an ideal habilitation model : - Writing a state of the art in habilitation models for identity and access management projects. - Writing functional and technical specifications. - Development - Unit testing - Documentation Technical environment : Java JEE (EJB3, JPA, JBoss AS), Flex 3, AS 3, BlazeDS, MySQL

Statistics, Econometrics, Machine-Learning, Deep-Learning, No-SQL, Spark, Hadoop, Data Visualization. In partnership with BNP Paribas, development of a bot able to successfuly pass the CFA exam (Charted Financial Analyst): Natural language processing / Deep learning.