GIZ ML
100% on-premise GPU clusters without external cloud dependencies.
RTX 5090 cluster, WanGP runtime, vLLM dynamic memory sleep/wake, 4GB disk LRU cache, and Whisper speaker diarization.
Architecture illustration
PRODUCT IN USE
Inspect the product in use.
Check the capture date and sample-data label, then select an image to inspect it. Operating performance, security requirements and customer outcomes are separate acceptance checks.
Giz operating settings and status
Enlarge view ↗
Giz operating settings and status
An administration view of operating settings and status. It is not evidence of air-gapped operation or security certification. 2026-09-13 · Sample data.
Explore this capability →Giz Computer browser terminal
Enlarge view ↗
Giz Computer browser terminal
The terminal tab and work area in Computer. Isolation, permissions and recovery are verified through separate execution checks. 2026-09-14 · Product screen.
Explore this capability →Confidential enterprise data and media pipelines cannot be trusted to external clouds
Transmitting sensitive engineering schematics, manufacturing inspection video feeds, unreleased product designs, and executive meeting recordings to third-party AI clouds (such as Runway, Midjourney, or OpenAI) introduces unacceptable enterprise vulnerabilities. Furthermore, escalating API usage fees erode operational profitability as volume scales.
Giz ML is a full-stack machine learning infrastructure powered by on-premise NVIDIA RTX 5090 clusters (ml-dmc1~9, mtl1~4) and the WanGP distributed worker runtime, natively serving ultra-high-definition video generation, real-time lipsync, 50-speaker diarization, and 32k-token document OCR entirely within the enterprise perimeter.
4 Engineering Pillars of Giz ML
| Engineering Domain | Giz ML Architecture | Operational & Business Value |
|---|---|---|
| RTX 5090 WanGP Fleet | NVFP4 quantization backend with NVIDIA VSR super-resolution and HEVC NVENC | Produces 60fps high-fidelity video completely on-premise with zero external API fees |
| vLLM Dynamic Memory Sleep/Wake | Instant VRAM reclamation and process sharing via /sleep?level=1 and /wake_up APIs |
Maximizes GPU hardware utilization by dynamically swapping between OCR and video workers |
| 4GB Disk LRU Cache | High-speed SQLite disk LRU (embeddings.sqlite3, 4,096 MiB) text encoder cache |
Eliminates text encoding compute latency (0ms overhead) across recurring prompt requests |
| High-Capacity Speaker Diarization | Combined faster-whisper large-v3-turbo and pyannote 3.1 pipeline |
Separates and transcribes up to 50 concurrent meeting participants with millisecond timestamps |
7 AI Media Pipelines Running In-House
- Video Generation & Frame Interpolation: LTX-Video foundation generation with ComfyUI RIFE 60fps interpolation.
- Real-time Digital Human Lipsync: MuseTalk and SoulX landmark synthesis from a single portrait and audio track.
- 50-Speaker Meeting Diarization: Multi-speaker discussion separation and timestamped automated transcript generation.
- 32k-Token Vision Document OCR: OvisOCR2 structured markdown extraction for multi-column papers, tables, and LaTeX.
- Speech Synthesis & BGM Composition: Expressive OmniVoice Korean/English TTS and Stable Audio 3 generation.
YOUR NEXT STEP
Bring us the workflow that needs to work.
Start with the files and systems you use today, and the work you want to change.
Ask Giz Agent
Explore capabilities and implementation scope using published product material.