Benoit Seguin

Simula (2024–Present)

Teaching AI models requires massive amounts of high-quality data, but generating it manually is impossible.

Simple automated prompting causes AI to create repetitive, low-quality data. I co-founded Simula with Hamza Harkous to solve this. We designed a programmable framework that independently controls data diversity, complexity, and quality without human intervention.

Our team scaled Simula into Google's primary internal multimodal data synthesis engine. It is now used by over 2,500 Googlers, was used to generate trillions of tokens, and achieved a 93% user satisfaction score, the highest among all data tools at Google. Used by hundreds of teams, today it is a key enabler for the Gemma ecosystem, provides the primary synthetic data backbone for Gemini safety classifiers, powers production user protection features like AI-powered scam detection, and extends much further across frontier AI applications and data use cases.

Highlights & Applications

• One of the core synthetic data generation engines for Google's open-weights Gemma 4 model family [Technical Report]
• Enabler for ShieldGemma 1 [Paper] and ShieldGemma 2 [Overview]
• Powers production features: Android Call Scam Detection [Pixel Drop] & Android Messages Spam Detection [Google Blog]
• Applications across FunctionGemma, Workspace Prompt Injection Defense, and Enterprise Security ML
• Research Paper (TMLR) • Google Research Blog Overview

Internal LLM Prompting & Agentic Bulk Inference Engine (2023–Present)

As LLM workflows evolved beyond simple chat prompts into complex agentic graphs and bulk synthetic generation, research teams needed a flexible way to manage prompt composition and run large-scale inference efficiently.

I created a lightweight, expressive prompt management toolkit and bulk inference engine that became popular among researchers across Google for rapid prototyping and large-scale batch generation.

Architecture & Highlights

• Expressive Prompt Templating: Extends modern templating engines with composable primitives for structured schema enforcement, multimodal context assembly, dynamic few-shot injection, and modular partials.
• Distributed Bulk Inference: Built horizontally scalable inference pipelines on top of Jax-on-beam, allowing complex multi-step and agentic generation workflows to execute across distributed accelerator clusters.
• Off-Peak Compute Efficiency: Leveraged opportunistic, off-peak datacenter capacity to run high-throughput batch inference workloads with significant infrastructure savings.
• Foundational Layer for Simula: Provided the core prompt rendering and batch inference substrate underlying Simula's data generation pipelines.

Scale & Impact

• Reached 10,000+ monthly active users, becoming a go-to toolkit across Google for structured prompt templating and bulk inference.
• Used to generate multi-trillion tokens across diverse research workflows and data generation pipelines.