Gilles L.

Manage batch and streaming AI / ML pipelines of billion products/day and ads/day on CPU / GPU using NLP / CV deep learning models for classification and unification tasks to enrich product catalogs and to improve ad safety.

• Owned a cross-business work that increased eligible ad campaigns by 3× by building a batch ML data

pipeline to unify 25B products: technical referent, defining PoC, roadmap, architecture, and coordinating 5 teams to

implement a hybrid architecture exact matching and vector search system KNN-based using FAISS.

• Led a technical strategy to improve scalability and data consistency with the introduction of Apache Iceberg as a new technology via a MVP on 2B products with peaks of 100K products/sec.

• Optimized infra costs to export cold data: DynamoDB → AWS S3 reducing annual costs by 54% (~$180K).

• Architected, optimized and scaled a critical component deployed globally into gRPC-based services, separating rule-based processing from ML inferences services across CPU / GPU infrastructure with monitoring: -50%

processing time and 4× throughput scale (500M → 2B products/day) with NVIDIA Triton.

• Led a taskforce of 6 members to deliver a top-priority e2e product feature for cross-retailers ad campaigns with a 9 months roadmap to unify 1.5B products (+16.3% the spends of the accounts).

• Founded a Hackathon team to build a vector search engine on e-commerce products (3rd price).

• Accelerated the enrichment flow of 500M products daily (hours to seconds) by building a ML online data pipeline instead of an offline one with Spark, Kafka and gRPC in Java/Scala.

• Taught courses for IASD Master on distributed processing and recommendation systems in Python (grades 4.6/5).

• Optimized the RAM of Spark jobs with -6TiB i.e -0.5% of available total RAM by rationalizing config and params.

• Supervised a ML engineer intern to enhance a classifier by using product image features.

• Automated data pipelines to create ML annotated datasets with a term search engine (+300M samples).

• Migrated a ML ecosystem to deep learning with transformer-based models using TensorFlow (7× speed-up in

processing time between PoC vs prod, +10% model accuracy).

• Refactored and upgraded search stacks ElasticSearch → OpenSearch (critical real-time 24/7 business services).

• Migrated Hadoop MR → Spark with mob programming from scratch (100K LoC → 50K LoC, x3 faster).

• Designed and implemented a research project to redefine user engagement metrics for large-scale e-commerce systems handling billions of events and users.