Organization Card

Innovius — Secure AI, built for trust

You use the AI. The AI does not use you.

Innovius builds the Secure AI Suite: private, governed AI for enterprises — on-premise, VPC or air-gapped — and the two products that run on it, ChtSafe for people and teams and AICNCT for connecting businesses to any AI. At the heart of all three sits ShinrAI (信頼 — Japanese for trust): a transparent, realtime semantic privacy layer for AI communication — PII detection, pseudonymization and lossless restore — shipped here as open-weight models, measured in the open, losses included.

🚀 ShinrAI 1.3 — the 15-locale generation

A joint development by Innovius × EECC Research Labs, trained on the JURECA supercomputer at the Jülich Supercomputing Centre.


🧠 Model	`shinrai-pii-m-v1.3` — PII detection across 15 locales (de en fr es it pl pt-BR pt-PT ru uk tr ar ja ko + he beta), 19 detection heads / 27 entity classes, 1024-token window, torch-free ONNX serving on CPU or GPU. Open weights under the Innovius Open License v1.0 — free for government, education, research and companies under $10M revenue; Apache 2.0 from 1 September 2028
🛡️ Try it live	`shinrai-pii-playground` — text, PDF, image and audio PII detection with replacement & lossless restore. Nothing you submit is stored
📊 Benchmarks	`shinrai-pii-benchmarks` — ShinrAI 1.1 / 1.2 / 1.3 and the field (Azure AI PII, Presidio, GLiNER, Liquid AI, OpenMed and more) in one harness on the same texts, every raw run file in `shinrai-pii-bench-results`

What ShinrAI does differently

Classical PII tools redact or hard-cut. A blanked text loses the information the AI needs to reason. ShinrAI replaces instead: every person, place, street and organization is swapped for a statistically faithful stand-in — matched in type and tier, cultural origin (51 classes), gender expression, name part and register — and passes six validation gates before it ships. The AI answers on natural text, and the reply is restored word for word on your side. The replacement map exists once, on your side; there is no server-side mapping store, by construction. The same models de-identify at scale: electronic patient files (ePA), clinical archives, banking correspondence, citizen case files.

Measured, not claimed

All ShinrAI numbers come from a sealed evaluation with acceptance gates frozen before measurement. Competitors run at their best threshold; ShinrAI runs at its default settings.

Suite (register named)	ShinrAI 1.3	Field
15-locale business & clinic letters, strict spans, median incl. Hebrew beta	95.0	Azure AI PII on the same texts: 53–76
Japanese letters, 300 texts (our name is Japanese, so is the test)	98.1	Liquid AI LFM2 PII Extract JP: 56.6
27 adversarial chat prompts (2026 privacy study), overlap F1	84.9	Azure preview 88.8 · Azure GA 81.1 · Presidio 75.6 · GLiNER2 74.3
ai4privacy everyday messages (de/en/fr/it), overlap F1	80.8–85.4	Liquid AI LFM2.5 PII 69.4–76.7 · GLiNER2 35.6–45.2 · the row leader is flagged for likely training overlap
Long-record redaction (REDACT, 1,000 records, 25 languages), partial F1	36.5	Azure 51.2 — our largest measured gap, the top 1.4 lever
Synthetic US paperwork (Gretel / Nemotron), partial F1	52.2 / 64.4	Liquid AI 84.5 / 87.3 — a paperwork-register track is scoped for 1.4

Wins and losses render identically on the board. No single number decides a PII stack — pick tools on measured numbers, ours included.

Model lineup

Generation	Locales	License	Status
`shinrai-pii-m-v1.3`	15 (he beta)	Innovius Open License v1.0, gated (automatic approval)	Current release — in production in ChtSafe and Secure AI Suite deployments
`shinrai-pii-m-v1.2`	6 (western)	Apache 2.0	Archive
`shinrai-pii-m-v1.1`	de · en · ja	Apache 2.0	Archive — the first open generation
ShinrAI 1.4	15 + Hebrew toward GA		In training at Jülich: long-document redaction, the paperwork register, all locales retrained alongside

Every version converts to Apache 2.0 within 24 months of release. Training data is synthetic — generated and cross-checked by openly released models, anchored in public registries and open data. Every record carries license provenance; no proprietary APIs anywhere in the pipeline.

👉 Join the model beta

Beta participants get early checkpoints matched to their hardware, and their language, domain and quantization votes shape the training roadmap. Follow this organization to see every release the moment it lands.

Where ShinrAI runs


ChtSafe	Private AI workspace for people and teams: 50+ frontier and open models behind ShinrAI-protected, onion-routed access, on browser, iOS, Android and an OpenAI-compatible API. Reusable AI coworkers (Team Members), customer personas (Sims) and media generation under one trust boundary
Secure AI Suite	The enterprise stack: Secure AI Engine (inference proxy with multi-tenant AAA, routing and cost controls), Secure AI Portal (RAG, agents, plugins), ShinrAI Encryption — on-premise, VPC or air-gapped, Helm and Docker
AICNCT	Connects your services to ChatGPT, Claude, Gemini, Copilot and Grok via MCP, llms.txt and universal APIs. Self-signup, instant access
ShinrAI Platform & Connector	Kubernetes-native engine with an OpenAI-compatible endpoint between your applications and any LLM; a desktop tray app that routes Claude Code, Codex and SDK tools through on-device protection

Who we are

Innovius UG is based in Teltow near Berlin, Germany. Our founding team has built deep-tech and security systems together since 2008: large-scale enterprise AI delivery from 2018, custom language models since GPT-2 for industry-grade document generation, and our own 24-node GPU cluster for BERT training and LLM fine-tuning. Innovius was incorporated in September 2024 to carry our core products, by then already three years in production, under one brand. In 2025 we launched ChtSafe and AICNCT; in 2026 we started training the ShinrAI models at Jülich and opened the public beta.

Today our enterprise customers are primarily SMEs and mid-market organizations across Germany and Austria. Innovius is a technology supplier to the KI-Navi Handel initiative — a KI.NRW flagship project by Handelsverband NRW and EECC bringing AI to retail businesses across North Rhine-Westphalia.

Partners & thanks

Developed together with the EECC Research Labs (EECC on Hugging Face) — our closest research partner and co-developer of the ShinrAI models, not a subcontractor. The EECC consulting team takes customers from the first assessment to a running deployment; VITAC, our technical implementation partner, brings ShinrAI onto endpoints and keeps it running.

Trained on the JURECA supercomputer at the Jülich Supercomputing Centre (Forschungszentrum Jülich), scaling onto JUPITER, Europe's first exascale system — supported by the WestAI initiative. Our deepest thanks to FZ Jülich and the JSC team: work like this is only possible because Europe's research infrastructure is open to it.

Training data is generated, cross-checked and evaluated entirely by openly released models — the Qwen (Alibaba), Gemma (Google DeepMind), Mistral and NVIDIA Nemotron families — on the base of mmBERT (Johns Hopkins CLSP) from the ModernBERT lineage, anchored in open data: GeoNames, Wikidata, OpenStreetMap, and the statistics offices and open-data portals of the countries we cover. Thank you for keeping frontier-quality open models and open data available.

Open weights, on principle

We fully support the Open Weights and American AI Leadership open letter published in July 2026 by an NVIDIA-led coalition of 50+ organizations: open weights are defensive assets — for security, for competition, for trust. The ShinrAI models are our contribution from Europe.