Where ideas are tested before they become products. Open source projects and experimental tools from our lab. We build in public and share our work with the developer community.
A research project: an Apertus 8B teacher model that learns to turn a capability definition into small specialist models (Apertus 0.5B–4B). It generates training data, finds where the student fails and teaches again. CSCS supports the training with GPU compute on Alps.
Small models can run offline on a laptop, a phone or an industrial gateway, but teaching one a narrow new task still takes lots of data and hand-built pipelines. We want to find out whether a model can learn to teach: diagnose why a small student fails, write targeted examples, and be judged by how much the student improves with how little data. Apertus is a natural base. It is fully open, and its 0.5B–4B models were themselves distilled from Apertus 8B. The name blends maravilla and maestra (Spanish for teacher).
One database for content, versioned like Git. It combines document storage, graph queries, vector search and full-text search behind a PostgreSQL wire protocol. Branching, time travel and ACID transactions give AI agents the isolated, reproducible state they need to reason and iterate.
AI agents and content-heavy apps usually mean a half-dozen services stitched together: Postgres for documents, Elasticsearch for search, a vector store for embeddings, blob storage for files, and custom plumbing for versioning. RaisinDB unifies all of that into a single Rust binary on RocksDB. Git-like branches give each agent task its own isolated state to mutate without corrupting shared memory, then merge back when done.
A tiny DSL for hand-drawn wireframes that lives inside fenced markdown code blocks. It has 83 element types covering desktop, mobile, AR, dashboards and AI/chat, plus framework-free web components and remark/markdown-it plugins to render them anywhere.
Wireframes usually live in design tools, disconnected from docs and code. Mermaid proved that fenced code blocks are a great home for diagrams, and we wanted the same for wireframes: a text-first DSL that diffs cleanly in git, renders in any markdown pipeline, and stays readable to humans and AI alike.
A self-hostable developer infrastructure platform combining git hosting, multi-tenant package registries (Docker/OCI, npm, Cargo, Maven), and containerized job execution in a single binary. Built for European organizations that need sovereign, auditable infrastructure.
European regulated sectors need developer infrastructure free from foreign government access laws. No existing solution combined git, all major package registries, and CI compute in a single self-hostable binary with full tenant isolation.
Svelte-inspired server-side Lua templating for Rust. Features component system, reactive syntax, live reload development server, and compiles templates to Lua for fast execution.
Edge-side personalization breaks SPAs: hydration expects the same content the client rendered, so changing things at the edge causes mismatches. Server-side templating solves this, but the developer experience feels dated. We wanted modern component-based DX for server templates, so we built this experiment to see what that could look like.
A modern, lightweight style guide generator for CSS/SCSS design systems. Supports W3C Design Tokens, programmable tokens with 46 build-time functions, fluid typography, and accessibility automation.
A project we've used internally for many years, now open sourced. Storybook felt too heavy for our needs. We wanted something lightweight, documented directly in the code. It also generates output that works well for both humans and AI tools.
Modern PHP runtime compiled to WebAssembly with instant cold starts and native performance. Features direct PHP-to-WASM compilation, reference counting GC, and a web-based playground.
PHP isn't used at the edge because of cold-start times from the traditional CGI/FPM model. We wanted to explore: what if PHP compiled natively to WASM? It is a learning exercise to test new WebAssembly spec features and build compiler expertise, and maybe one day it becomes a full PHP-to-WASM compiler.
A lightweight service worker library for efficient caching strategies. Pattern-based URL matching, support for GET and POST requests, configurable expiration, and IndexedDB storage for offline access.
Legacy and server-side apps need offline-first capabilities without a full rewrite. Zephyr is a drop-in service worker solution with deterministic, pattern-based caching. It treats IndexedDB as a first-class citizen for apps that need reliable offline storage.
All our projects are open source. We welcome contributions, feedback, and ideas from the community.
View on GitHub
