.png)
Synthetic Data for AI agents
We simulate expert-level reasoning and domain-specific processes to generate training data that builds true specialization into your models. Synth handles the cold start problem, covers the long tail of edge cases through engineered simulations, and runs entirely on-premise to keep sensitive data under your control.
Learn more
Fully Open Data for AI
We curate and structure the world's largest rights-cleared and provenance-based dataset for LLMs - government records, legal archives, scientific literature, multilingual sources - so you can plug it directly into your models, RAG pipelines, and MCP servers.
Learn more
AI-Native Tooling for Agents
Turn your messy siloed documents into a single, structured, compliant data asset for your agentic AI workflows. Your AI systems and agents get not only the right information but rich trustworthy context - higher accuracy on a wider range of processes. A built-in privacy firewall handles PII before anything leaves the secure zone. Deployable fully on-premise.
Learn more
BlogMultilingual synthetic data is now judged almost entirely by other AI models. Yet we rarely know whether these judges read Latvian or Italian as well as they read English. The problem is circular. Models trained mostly on English tend to write text that sounds translated, and the judges asked to catch it inherit the same bias, while human experts cannot audit at that scale and a second model may share the first one’s blind spots. So who judges the judges?
BlogIn partnership with AI-BRIDGES at the Digital Humanities Research Hub, University of London, we are releasing an open recursive harness in which language models search Wikidata through code, instead of pasting graph results into their context. In it, Qwen3.8-27B, an open-weight model served on a single GPU, answers 74 of 100 test questions, the best score in our evaluation, and gpt-6-luna doubles its accuracy on multi-hop questions. Alongside it, we release the 10,235 search traces recorded in the harness, for training smaller models to do the same.
.jpeg)
BlogOn the occasion of the AI for Good Global Summit in Geneva, we are open-sourcing the models, datasets, and deployment frameworks presented in this post: the reasoning models framework for RAG, our on-device benchmarking harness for cache-augmented generation, and the French-language retrieval pipeline for Raspberry Pi and Android. Everything described here runs fully offline, on hardware costing less than €100.
.png)
Use CasePleias trained a 600-million-parameter specialized model for RATP to detect and interpret safety signals in Parisian Subway users’ messages - combining a fully synthetic training pipeline and designed for on-premise deployment. After only three months of development, the model, beating closed models 200x times its size, is now in production at RATP’s sovereign infrastructure on Scaleway.
Use CaseResearchers built an AI medical assistant that works without internet for health workers in rural West Africa. It runs on old Android phones and handles local languages to give workers quick treatment guidance.
Use CasePleias and SpineDAO are partnering to build AI systems that safely scale expert spine care for back pain, the world's leading cause of disability. The project tests whether small, structured-reasoning models can outperform large generic LLMs in high-stakes clinical settings.
