gateway 01 
★ flagship innovation · 2025 · Live Corvux
The Microverse Battery (for your brain)
A local-first 'second brain'. Your notes are plain Markdown files in a roost on your own machine — no cloud lock-in. Corvux indexes them into a live [[wiki-link]] knowledge graph you can traverse: backlinks, neighbors, shortest paths, and PageRank to surface your most central notes.
- ◆Local-first Markdown roost with sub-second re-indexing on file changes.
- ◆Knowledge-graph engine — backlinks, forward links, N-hop neighbors, path-finding, PageRank hubs.
- ◆Tasks, reminders (OS notifications) and properties/tags layered over plain notes.
- ◆Ships an MCP server so AI agents can read, write and reason over your roost directly.
gateway 02 
★ flagship innovation · 2026 · Active build Portty
Your phone is a portal gun for terminals
An open-source, zero-config terminal to your own machines from anywhere. Portty keeps shell sessions alive and lets you switch between them from your phone over an encrypted iroh P2P tunnel.
- ◆Persistent, switchable PTY sessions with scrollback and high-throughput binary streaming.
- ◆Secure PIN pairing and token-based fast resumption/reconnect over encrypted iroh P2P tunnels.
- ◆Adopt any terminal mid-work with the `portty share` ConPTY/named-pipe relay.
- ◆Everyday shell-profile hook that makes all terminal sessions portal-ready from birth.
- ◆Custom mobile UI with accessory key bar, Ctrl-key latching, and safe-area compatibility.
Rustiroh / QUICTauri 2SolidJSxterm.jsConPTYE2E Encryption
Digital Avatar
A talking interdimensional assistant
LLM-based conversational system that understands queries, retrieves context and responds — wired to Riva ASR/TTS for full speech I/O. Fine-tuned with PEFT/LoRA/QLoRA, guardrailed to its domain, deployed on Triton.
LLMRAGRiva ASR/TTSLoRA/QLoRAGuardrails
Credit Engine
Knowledge graphs that decide
End-to-end knowledge-graph pipeline over unstructured text. Extracted entities with NeMo LLMs, built the graph with NetworkX/PyViz, ran PageRank & community detection, then injected the graph back into an LLM to predict the next best action.
Knowledge GraphNeMoNetworkXGraph Mining
GPT-20B Pre-training
Multi-GPU, multi-dimension
Pre-trained a 20-billion-parameter GPT model end-to-end using DeepSpeed pipeline parallelism on an NVIDIA DGX cluster.
DeepSpeedPipeline ParallelDGXDistributed Training
Kantar Vision
See it, read it, identify it
Multi-model pipeline for text + image extraction and food-item identification. os2d one-shot learning, YOLOv7 brand detection, OCR+NLP fusion, plus a face-analysis model for age/gender/region attributes.
os2dYOLOv7OCRFace Analysis
NVIDIA Launchpad Labs
Three labs, one toolkit
Automatic close-captioning (NeMo/TAO/Riva TTS), vehicle damage estimation (TAO object detection on Triton with dynamic batching), and demand forecasting (RAPIDS + Dask-XGBoost, Plotly Dash dashboards).
TAORAPIDSDaskTritonPlotly Dash
