Etienne Boisseau

Development and industrialization of AI and NLP solutions within the entire BNP Paribas group as an AI Engineer in a dedicated team to advise group entities on their technical and strategic decisions in the field. Lead maintainer and designer for one and a half years of a Python framework for scaling RAG initiatives in the group, used by about fifteen teams within the group.

Building and maintaining CI pipelines, Python libraries for several projects and entities in the group.

Strong focus on building automated evaluation pipelines to ensure the reliability of LLM agents, RAG assistants, search systems, and PDF parsers.

Vignoblexport wanted to implement an incident detection algorithm but could not tolerate any false negatives. This required building an extremely robust processing chain and finely controlling data quality.

In addition to detecting incidents, the company wanted to predict them before they occurred by leveraging its 10 years of historical data.

Tasks performed:

- Needs assessment, creation of specifications and a provisional schedule.

- Investigation of the source of the problem with the existing algorithm: data quality issues.

- Creation of a database schema to address the problem (6 new tables needed to be created).

- Creation of a Python package to efficiently manage the new concepts introduced, populate, and continuously update the database.

- Design of 5 atomic workers fulfilling different roles related to data sourcing, processing, team alerting, and monitoring of the entire processing chain. The separation of tasks allowed for low coupling between workers and maximized desirable properties such as idempotency and statelessness of operations.

- Implementation of 5 Python clients using the previously created package, implementing the 5 designed workers.

- Creation of datasets for training predictive models.

- Feature engineering, feature selection, and model selection for prediction tasks.

- Creation and exposure of an API to query predictive models.

- Deployment of data processing and prediction workers using Docker.

Faced with very high inference costs for its deep learning models, the client wanted to reduce this cost without compromising quality.

By using cutting-edge neural network optimization methods (quantization, distillation, pruning, graph optimization) and optimizing the use of the client's cloud resources, we reduced the inference cost of large NLP models by a factor of 6.

Tasks performed:

- Needs assessment, creation of specifications and a provisional schedule.

- Creation of a benchmarking tool to evaluate model performance and cost under different optimization parameters and backends.

- Conducting a comprehensive study establishing optimal choices for model optimization based on various criteria. Writing a report detailing these conclusions.

- Deployment of the best solution found, resulting in 85% savings on inference costs.