NEUROGLYPH FORGE

What we are building: English keystroke-aligned EEG from your EMOTIV EPOC X, fine-tuned decoders on your RTX 4070, reasoning in Hermes Agent, generative assets via Comfy Cloud MCP, and scene control in Unreal Engine MCP — the sensory-motor front of MindBot / synergetic cognition.

GitHub HF Model HF Dataset Brain map Architecture (MD)
NeuroGlyph banner

End-to-end pipeline

1 · RECORDCortex Premium EEG + QWERTY key times
→
2 · EPOCHS−200…+300 ms · 14×25 @ 50 Hz
→
3 · TRAINTinyB2Q / EEGNet · hand → zone → intent
→
4 · HERMESMCP tools · confidence gates
→
5 · ACTUnreal BP events · Comfy visuals

Not Meta-lab 64ch Brain2Qwerty weights — your 14-channel EPOC geometry, English-first labels, deployable on consumer GPU.

Brain test data (what we use today)

Synthetic (CI / smoke)

Generated 14ch epochs for pipeline tests. scripts/download_public_datasets.py synthetic

Gear test (mock Cortex)

Mock WSS CSV + injected key events — proves raw → processed → fine-tune on your PC. scripts/test_gear_finetune.py

Staging (ST) / HF dataset

Raw sessions staged for upload: neuroglyph-epoc-typing-en-v1. Catalog: data/catalog.json

Gold path (your headset)

~10 min English typing session → data/raw/ → fine-tune with frozen encoder. docs/FIRST_SESSION_PROTOCOL.md

SourcePurposeOn HF?
Synthetic handTrain smoke, adaptive engineNo (regen locally)
Gear / mock sessionsPipeline regressionYes (dataset repo)
Live EPOC (future)Personal decoder weightsYou choose (push script)

Pre-download model & fine-tune on your machine

Published checkpoint (hand, English path):

python scripts/download_hf_artifacts.py
# → checkpoints/hf/NeuroGlyph-EPOC-Typing-v1/
# → data/hf/neuroglyph-epoc-typing-en-v1/

python scripts/run_swarm_pipeline.py   # full YOLO: data → train → fine-tune → smoke

Local GPU: RTX 4070 12GB + PyTorch cu124 (docs/LOCAL_GPU.md). Inference can stay CPU; training uses CUDA when available.

Hermes Agent (in-repo MCP)

NeuroGlyph MCP server

  • preprocess_session — raw → epochs
  • train_decoder — hand / zone / intent
  • run_live_decoder — gated predictions
  • send_to_unreal — bridge to UE

Adaptive situations

neuroglyph_agent/adaptive_engine.py reads signals (live vs synthetic, checkpoints, queue) and writes data/adaptive_brief.md so the swarm pivots without re-steering.

Comfy UI / Comfy Cloud MCP

Role in the stack: generative layer for promo art, brain-map visuals, stream overlays, and experiment thumbnails — not the EEG decoder itself. Hermes can call Comfy Cloud MCP (workflows, templates, image gen) while NeuroGlyph handles BCI.

  • Brand assets under assets/brand/ (GPT-Image-2 pipeline)
  • Live stream: BCI intent → Hermes narrative → Comfy for visuals → OBS / reality live
  • Keep API keys in Hermes auth — never commit to this repo

Unreal Engine MCP

Editor HTTP MCP with Mcp-Session-Id: list_toolsets, call_tool. Predictions map through neuroglyph_unreal.ue_actions.prediction_to_unreal_action (e.g. MoveLeft, SpawnDebugActor) when confidence ≥ 0.75.

Setup: docs/UNREAL_SETUP.md · skill: unreal-engine-mcp in Hermes.

MindBot / synergetic cognition

NeuroGlyph emits PredictionEvent JSON into the MindBot bus — Hermes plans, dreams/CoT can comment on intent, Unreal executes. Export trajectories (context + plan + outcome), not raw EEG dumps, for MindBot training.

MINDBOT_INTEGRATION.md

Quick start

git clone https://github.com/TheMindExpansionNetwork/neuroglyph-forge.git
cd neuroglyph-forge
python -m venv .venv && .venv/Scripts/activate
pip install -e ".[cloud,recorder]"
pytest tests/ -q
python scripts/download_hf_artifacts.py
python scripts/run_swarm_pipeline.py