harciufff - Stardance

Ship Pending review

### What I Made
I built "Cena", an event-driven, edge-native system for real-time environment understanding. It runs a multi-threaded architecture that combines geometric telemetry (YOLO-World + MediaPipe) with asynchronous semantic descriptions (Moondream-2), logging everything locally into Markdown files inside Obsidian without vector database latency.

### The Challenge
Managing heavy AI models locally on an edge device without performance throttling or memory leaks. I had to design a loosely-coupled asynchronous pipeline where the semantic layer runs on a background thread every 10 seconds, ensuring the geometric layer can maintain a smooth ~30 FPS with GPU Metal/MPS acceleration on my Mac M4.

### What I'm Proud Of
The zero-indexing latency architecture. Instead of relying on heavy cloud vector databases, integrating the output directly into Obsidian markdown files creates a lightning-fast, local-first memory pipeline inspired by the LLM-as-an-OS concept.

### How to Test It
Clone the GitHub repository, set up a Python virtual environment, and install the requirements. You will need a local camera feed and an Obsidian vault path configured in your script. Run `python3 main.py` to start the pipeline and press 'q' to safely shut down.

1 devlog
6h

Try project → See source code →

Open comments for this post

@harciufff on cena · 23 days ago

5h 56m 54s logged

Hey everyone! I’m building “Cena”, a local-first visual copilot system that provides real-time geometric telemetry and asynchronous semantic understanding.

Note for the Shipwrights: I actually spent around 35-40 hours programming the core architecture and multi-threading structure last week before finding out about Stardance and installing the Hackatime extension. You can verify my daily development dates and progress through my GitHub commit history!

How the system works:

Geometric Layer (~30 FPS): Uses OpenCV to capture frames, feeding them to YOLO-World and MediaPipe Hands (running with GPU Metal/MPS acceleration on my Mac M4).
Semantic Layer: A background thread invokes Moondream-2 every 10 seconds to get a text description of the scene.
Agent Layer: As shown in the screenshot, the engine merges everything into plain markdown files inside Obsidian (LiveContext.md and PersistentContext.md). Inspired by Karpathy’s LLM-as-an-OS concept, this avoids vector database indexing latency on edge devices.

The code is currently split into two versions (“Cena” and “CenaAmd”) representing the performance updates and VLM optimization of the pipeline. Check out the site here: https://cenadev.vercel.app/

Check out the full repository here: https://github.com/harciufff/cena