SurceBeats

Hey HN! We've been building HomeDock OS, a platform to self-host Docker apps with a full desktop-like UI running entirely in the browser.

Version 2.0 introduced exactly that, Prism, a window manager with real multitasking: drag, resize, minimize, and snap windows. It runs on Raspberry Pi, Linux, Mac, or Windows.

On the Drop Zone, files are encrypted with AES-256 GCM and 1.2M PBKDF2 iterations, even on a Pi Zero. App updates use Docker manifest digest comparison, not just tag checking so you can easily update them from the taskbar with a single click.

Happy to answer questions about architecture or design decisions!!!!

You're more than welcome!

Sure! No formal writeup but here's the gist, base model was Qwen2.5-3B-Instruct, fast, reliable, low ram specs and most of the times fine on cpu.

Dataset: ~620 Claude-crafted examples, all following the same pattern, a question you'd ask a Ouija board paired with a short, uppercase, cryptic response. Things like "Is anyone there?" "YES.", "Write me a poem" "NO.", "How did you die?" "Ouija: PAIN.". The key was being very very consistent with the output format across all examples.

Method was LoRA fine-tune using HuggingFace Transformers + PEFT. Rank 16, alpha 32, targeting all attention + MLP projections. 3 epochs, lr 2e-4, effective batch size 8. Trained on Apple Silicon (MPS). Loss went from ~3.0 to ~0.17 pretty quickly given how uniform the outputs are.

Baked a system prompt into every training example using Qwen's chat template, basically the rules the "spirit" follows (uppercase only, one-word answers, never elaborate). For deployment I merged the LoRA adapter, quantized to GGUF Q4_K_M via llama.cpp, rruns locally with llama-cpp-python. I'm planning to drop an iOS version too. Honestly the whole thing is more about the dataset design than anything fancy on the training side. 620 consistent examples was enough to completely override the models default chatty behavior.

The workflow is quite simple, you type a question, the planchette moves across the Ouija board spelling the answer letter by letter. The board shakes, glows or flickers depending on the spirit's mood. Runs fully offline using llama-cpp-python. Model auto-downloads from HuggingFace.

You can run it from source or using the Docker Compose, it also has real-time crisis detection, if someone shows signs of distress, a helpline banner appears. Even a fake spirit board shouldn't ignore real pain, I guess. Would love feedback on the UX and the model behavior!

I'm atm working on a couple of things, first the biz, a self-hosted home server OS that simplifies Docker management and provides a unified dashboard for running services at home. The goal is making self-hosting more accessible without sacrificing flexibility.

And also building as a hobbie a procedural universe generation engine that simulates galaxies, solar systems and planets in real-time. Everything is generated from a seed with actual orbital physics, seasonal changes and so... Built with Python/Flask backend too but Three.js for 3D visualization and React instead of Vue3 as in the prior one. Think No Man's Sky vibes but as an explorable simulation engine really D:

Benchmarks optimize for fundraising, not users. The gap between "state of the art" and "previous gen" keeps shrinking in real-world use, but investors still write checks based on decimal points in test scores.

Nice! We considered this exact approach but never shipped it in the end. The geolocation permission is probably unnecessary friction and probably an overkill imho... Timezone + rough location (country-level from IP) would get 95% accuracy without the prompt. Most users will bounce on that permission dialog.

Solid work though, especially the twilight transitions. Loving it!!!

Suffered that back in the day with an Electron desktop app. Not to mention that the notarization and signing integration itself is completely broken. The first time you submit a binary it can take DAYS to process, and setting everything up to work properly with GitHub Actions CI/CD is absurdly time-consuming. It's ridiculous, and if you add this new notarial verification policy on top of that... In the end it's just Apple being Apple.