Jean-Marc Soumet

Senior Machine Learning Engineer, Applied Research

Applied Research team building the Adobe Express Assistant, an LLM-powered agent that lets users create and edit designs through natural language.

• Build and ship core product code for the Adobe Express Assistant, integrating LLMs and Vision-Language Models (VLMs) into design creation and editing workflows.
• Design evaluation frameworks for assistant quality, including LLM/VLM-as-judge pipelines that automatically score model responses and rendered designs at scale.
• Develop synthetic data generation pipelines that produce training and evaluation datasets for LLMs and VLMs.
• Help set up training and serving pipelines for LLMs and VLMs, including fine-tuning and deploying custom first-party (1P) models.

Principal Machine Learning Engineer / Manager

Built Ironclad's machine learning function from the ground up: shipped the company's first NLP models to production, then grew and led a cross-functional team of ML and full-stack engineers.

• Architected Ironclad's ML platform end to end: data preprocessing, PyTorch model training on NVIDIA GPUs in Google Cloud (GCP), and scalable prediction serving.
• Led development of a Retrieval-Augmented Generation (RAG) system combining LLMs, VLMs, and vector databases to extract and retrieve information from millions of customer PDFs.
• Shipped BERT-based transformer models for text classification and Named Entity Recognition (NER), and integrated LLMs (GPT, Gemini, Claude) for advanced information extraction.
• Designed zero-shot classification techniques that accurately classify contract clauses with no model training required.
• Managed four ML engineers and two full-stack engineers, owning hiring, mentoring, and onboarding; shaped the company's AI strategy with customers and senior leadership.

Senior Machine Learning Engineer

Led machine learning for the voice command team, owning the speech-to-text and natural language understanding (NLU) backend that turns spoken queries into on-device intents for content navigation.

• Owned the end-to-end speech-to-text and NLU backend, processing natural language queries from consumer devices at scale.
• Built data and training pipelines, including synthetic data generation tools for multiple languages and markets.
• Optimized the system for low-latency voice command responses, critical to the on-device user experience.
• Ran data processing and model training on AWS (S3, ECS, EC2, SageMaker), orchestrated with Airflow.

Progressed from Senior Software Engineer to Principal Machine Learning Engineer across four teams, spanning big data analytics, ML platforms, and voice AI.

• Led backend development for the Einstein Voice Assistant app, letting customers query and update Salesforce data by voice.
• Trained and deployed deep learning models for intent matching and entity recognition, and built the NLP and Automatic Speech Recognition (ASR) APIs.

• Designed scalable microservices for dataset uploads, training job scheduling, and real-time inference.
• Built model templates and fine-tuning scripts for LSTM, BERT, and Faster R-CNN architectures.
• Wrote high-performance C++ JNI adapters for TensorFlow and PyTorch models, speeding up inference and simplifying integration with Java services.

• Led a team of data engineers building ETL pipelines that moved terabytes of data daily from Salesforce's global data centers to AWS S3.
• Architected a Scala/Spark platform for batch processing of customer data, powering AutoML inference across Service Cloud, Sales Cloud, and Marketing Cloud.

• Built user behavior analytics for the SalesforceIQ product with Kafka, Flink, Spark, Hadoop, SQL, and S3.
• Managed a petabyte-scale Postgres data warehouse, including schema design, data extraction, and processing.