Head of Blablador — the open-model LLM service of Forschungszentrum Jülich and the Helmholtz community. Jülich Supercomputing Centre (JSC) · Jülich, Germany
I lead Blablador, the service that gives Helmholtz researchers an OpenAI-compatible endpoint to models we run ourselves: on JUWELS Booster, on haicluster, and on the federation sites. My work is the whole stack around it — serving large open models with vLLM, getting big models onto GPUs that were not built for them, the routers and monitoring that keep them reachable, the native apps people actually chat in, and the documentation.
Most of what you find here grew out of running that service, plus a long-running pile of small hardware projects.
- ✉️ [email protected]
- 🌍 surak.eti.br · surak.github.io · LinkedIn
- 🧠 Blablador: web interface · API documentation
- 📱 Blablador app: Mac, iPhone, iPad, Android, Windows, Linux, Raspberry Pi
| Project | What it is |
|---|---|
| blablador-ios | The native Blablador app: Mac, iPhone and iPad in SwiftUI, Android in Kotlin, Windows/Linux/Raspberry Pi in Rust. Not in the App Store or Google Play — iPhone and iPad builds come from source in Xcode, Android installs the developer-signed APK from the releases, the rest are release archives. (landing page) |
| homebrew-tap | brew install --cask surak/tap/blablador |
| depth-map-through-vllm-router | VGGT-1B 3D mapping behind an OpenAI-compatible API, so it can sit next to the vLLM models behind the router: a robot car maps a room by taking photos and sending commands, and photos of a place become a top-down map and a coloured point cloud. (demo) |
| blablador-paper | The paper: the LLM service of Forschungszentrum Jülich |
| decision-model-frontend | The decision examples behind the Clef decision model — give a state a set of typed questions, get probabilities instead of prose |
| semantic-routing-llm-plugin | An Apache APISIX plugin that lets a System One model (Clef) score which LLM instance a request belongs on |
| Project | What it is |
|---|---|
| HyperQwen | Large Qwen models, fast, on small cards: Qwen3.8-27B on a single 24 GB GPU — 127 tok/s single-user, ~1,035 tok/s at 64 concurrent, 150k–262k context. vLLM patches, a requantisation pipeline and benchmarks |
| HyperArete | Calibrated W4A16 Arete-27B weights and a reproducible HyperQwen serving recipe for RTX 3090 GPUs |
- diskInventory-NG — see where your disk space went: a fast cushion-treemap disk usage visualizer for macOS, in Rust (a tribute to Disk Inventory X)
- tech-news-digest — automated tech-news digest: 151 sources, one pipeline, a daily AI/agent/crypto/frontier-tech digest
- qwen3-tts-cli — text to speech with Qwen3-TTS on macOS, from the command line
- macos-root-ca-cleanup — remove two root CAs from Apple's OpenSSL trust bundle, with backup, validation and revert
- arducam-web-control-v2 — web and CLI control for an Arducam 16 MP autofocus camera on a pan-tilt platform, driven by a Raspberry Pi 5
- brickremote — drive Lego-compatible Bluetooth motor bricks from a computer or tablet
- sun-tilt-ios — iOS app for the optimal solar panel angle
I teach the JSC course Bringing Deep Learning Workloads to JSC Supercomputers and give Blablador talks and workshops for Helmholtz centres, universities and research organisations — what the service is, how to put your own model on it, and how to reach it from code, editors and agents.



