Mission Control UI

System Architecture & Local Execution

While the moro CLI provides complete automation for scripting and CI/CD pipelines, complex fine-tuning operations benefit immensely from visual, real-time feedback. MoroAI provides Mission Control: a self-contained web application served entirely from the local Python package.

When launched with moro dashboard, the system starts an embedded, high-performance ASGI server (Uvicorn + FastAPI) and binds exclusively to localhost (or a user-specified private subnet interface). No telemetry or tracking scripts are loaded; assets are served locally from the package distribution.

$ moro dashboard --port 3000 --host 127.0.0.1

╭──────────────────────────────────────────────────╮
│   MoroAI Mission Control Dashboard v0.1.0        │
│   Web Interface:  http://localhost:3000          │
│   REST API:       http://localhost:3000/api/v1   │
│   WebSocket:      ws://localhost:3000/api/v1/ws  │
│   Storage:        SQLite (.moro/moro.db)         │
╰──────────────────────────────────────────────────╯
[READY] Press Ctrl+C to stop the dashboard server.
    

Real-Time Streaming Engine (SSE & WebSockets)

Standard machine learning dashboards rely on aggressive REST polling (e.g. fetching /status every second), which introduces latency and CPU thrashing during heavy backpropagation.

MoroAI Mission Control implements a dual-channel pub/sub telemetry pipe:

• Server-Sent Events (SSE) (/api/v1/stream): Delivers unidirectional, lightweight 60Hz telemetry ticks for active training step, rolling moving-average loss, and exact VRAM allocations directly to client SVG/Canvas charts.
• Bi-Directional WebSocket (/api/v1/ws): Facilitates instant command dispatching (such as dynamic learning rate nudges, emergency pause, or manual checkpoint creation) with sub-millisecond execution roundtrips.

Core Modules in Mission Control

1. Interactive Dataset Studio

Drag and drop raw .jsonl, .csv, or .parquet files directly into the browser. The studio immediately runs the Epistemic Data Compiler in a background worker process, rendering:

• Interactive token length distribution histograms
• Positive Pointwise Mutual Information (PPMI) density curves
• Visual identification of protected MI Guard Landmark samples versus pruned boilerplate

2. Pre-Flight VRAM Simulator

Before starting training, operators can adjust sequence length (e.g. 1024, 2048, 4096) and LoRA rank ($r=8, 16, 32, 64$). A dynamic gauge continuously predicts peak VRAM against available GPU memory and illuminates safety thresholds (Green: Safe, Yellow: Tight, Red: OOM Inevitable).

3. Evaluation Comparator & Completion Diffing

Compare base model versus fine-tuned checkpoint completions side by side on standardized enterprise test sets. Responses are diffed token-by-token with color-coded highlighting of strict rule compliance, hallucination markers, and Wasserstein distribution drift scores.

4. One-Click Runtime Deployment

Once an evaluation gate is certified, the operator can click Deploy to Ollama. Mission Control invokes local quantization (NF4/GGUF), compiles the Modelfile, and registers the model in the local Ollama daemon without leaving the browser.

REST & WebSocket Integration

For programmatic monitoring or embedding into internal corporate portals, Mission Control exposes comprehensive endpoints: