NemoStage

Live presentation co-pilot for high-stakes technical demos.

NemoStage turns a static PowerPoint deck into an adaptive, AI-powered presentation assistant. It watches what you say in real time, detects when you go off-script, and silently generates supplemental slides to support the detour — all on local hardware.

---

What it does

Capability	How
Slide rendering	Parses PPTX XML with Docling; renders text, images, and tables on a pixel-accurate canvas
Off-script detection	Embeds slide text with SentenceTransformers; classifies live transcript against the index
Supplemental slide generation	Fires a background agent (Gemma4 26b via NemoClaw) when coverage status = not_covered
Audience Q&A	Audience scans QR → submits question → agent answers with deck context → broadcast to presenter
Material RAG	Upload supporting docs (DOCX/PDF/TXT) alongside the deck; answers draw on both

---

Architecture at a glance

Laptop (Electron/React)
  └── nemostageApi.ts ─── HTTP ──► FastAPI backend (DGX Spark, port 8000)
                                      ├── ChromaDB  (vector index, local)
                                      ├── Brev Ollama pool  (nemotron-3-nano:4b × 4)
                                      └── OpenClaw gateway (ws://127.0.0.1:18790)
                                            └── NemoClaw sandbox
                                                  └── Gemma4 26b (main / audience / slidegen)

---

Hardware

• Frontend: Any laptop (macOS/Windows/Linux)
• Backend: ASUS Ascent GX10 (DGX Spark) — Blackwell GPU, local inference
• Network: ucscguest Wi-Fi or Tailscale VPN

---

Quick start

Frontend (laptop)

cd NemoStage
npm install
npm run dev          # hot-reload dev build
npm run build:mac    # production installer

Backend (DGX Spark)

source /home/asus/nemostage-venv/bin/activate
cd /home/asus
uvicorn nemostage-server:app --host 0.0.0.0 --port 8000

# or via systemd
sudo systemctl restart nemostage
sudo journalctl -u nemostage -f

Deploy changes from local:

# SFTP server.py to /home/asus/nemostage-server.py, then:
sudo systemctl restart nemostage

---

Key endpoints

Method	Path	Purpose
GET	/status	Health check
POST	/sandbox/uploadpptx	Upload PPTX, build vector index
POST	/sandbox/presentation-materials	Upload supporting docs
POST	/presentation/start	Create session
POST	/presentation/transcript	Analyze live transcript chunk
GET	/presentation/{id}/generated-slides	Poll for generated slides
POST	/audience/question	Audience submits question
GET	/audience/qa/recent	Fetch answered Q&A pairs
WS	/ws/presenter	Real-time Q&A broadcast

---

Tech stack

Layer	Technology
Desktop shell	Electron 39
UI	React 19 + TypeScript 5
Build	electron-vite + Vite 6
State	Zustand 5
Backend	FastAPI (Python 3.11)
Embeddings	sentence-transformers/all-MiniLM-L6-v2
Vector DB	ChromaDB (persistent, local)
LLM inference	Brev Ollama (nemotron-3-nano:4b), NemoClaw (Gemma4 26b)
Agent runtime	NemoClaw / OpenShell sandbox
Slide parsing	Docling + custom PPTX XML parser

---

Project layout

NemoStage/                    # Electron app
  src/main/                   # Main process, IPC, media protocol
  src/preload/                # Context bridge
  renderer/src/               # React UI
    components/               # 13 UI components
    services/nemostageApi.ts  # Backend HTTP client
    store/presentationStore.ts # Zustand state
nemostage_backend/
  server.py                   # FastAPI server (2300 lines)
  slide_templates/            # Curated template library
backend/transcript/           # Legacy Whisper.cpp integration
PROJECT.md                    # Vision & roadmap
AGENTS.md                     # NemoClaw agent reference
ARCHITECTURE.md               # Full system architecture