Industrial AI for SMEs
We believe industrial AI should be four things: secure, explainable, practical, and affordable for SMEs. Anything else is a demo — not a system a factory can rely on.
Most generative AI for industry assumes you will ship operational data to a hyperscaler and pay per token forever. Small and mid-sized manufacturers cannot — and should not — accept that trade. Production data is the business.
Sakaki packages a local LLM, a multi-agent workflow, a Text-to-SQL engine and an enterprise knowledge base into an on-premise appliance. AI as infrastructure, not as a SaaS dependency — no OpenAI, no cloud, no data egress.
Secure
Every model runs inside your infrastructure. No cloud calls, no telemetry of production data, GDPR-compliant by construction.
Explainable
Answers ship with the SQL, the sources and the knowledge base excerpts they came from. Grounded outputs over clever ones.
Practical
Purpose-built for manufacturing terminology, MES workflows and shop-floor decisions — not a generic chatbot dressed up for industry.
Affordable
One-time appliance, no per-token cloud bill. Priced and shaped for SMEs, machine builders and automation integrators.
Why your data never leaves the plant.
Straight answers about how Sakaki runs locally, why we don't call OpenAI, and what "air-gapped" actually means on the shop floor.
[01]Do you use OpenAI, Anthropic or any external cloud LLM?
+
Do you use OpenAI, Anthropic or any external cloud LLM?
No. Sakaki runs open-weight local LLMs (Qwen, Llama and Text-to-SQL specialists) entirely inside the appliance. No API calls to OpenAI, Anthropic, Azure OpenAI or any hyperscaler — ever. Your prompts, schema, SQL and answers never leave the box.
[02]Where does my production data physically live?
+
Where does my production data physically live?
On your appliance, on your network. The Text-to-SQL agent reads a near-real-time sandbox mirror of your SQL Server / Postgres / MySQL / Oracle. Nothing is uploaded, no telemetry of production data is sent home, and the appliance can be fully air-gapped from the internet.
[03]What does "air-gapped" actually mean here?
+
What does "air-gapped" actually mean here?
Two layers. (1) Network: the appliance can operate with zero WAN egress — inference, RAG and reporting all happen on the LAN. (2) Data: agents only read a physically isolated sandbox copy of your production database, so there is zero risk of write-back to the live MES.
[04]Can it work fully offline?
+
Can it work fully offline?
Yes. Once deployed, the appliance requires no internet connection to answer questions, generate reports or run predictive models. Updates are shipped as signed offline packages that you apply on your schedule.
[05]How is this GDPR and IP-compliant by construction?
+
How is this GDPR and IP-compliant by construction?
Because no personal, operational or engineering data ever leaves your infrastructure, there is no cross-border transfer, no third-party processor and no shadow training on your data. Model weights are static and local; your data is never used to train anyone's model.
[06]Who can see what? How is access controlled?
+
Who can see what? How is access controlled?
The appliance integrates with your existing identity provider (AD / LDAP / SSO). Agent actions are logged with the user, the prompt, the generated SQL and the sources used — every answer is auditable and reproducible.
[07]Why not just use ChatGPT Enterprise or an Azure OpenAI deployment?
+
Why not just use ChatGPT Enterprise or an Azure OpenAI deployment?
Cloud LLMs still ship your schema, prompts and results outside your perimeter, charge per token forever, and depend on WAN uptime. Industrial sites need deterministic latency, zero data egress and a fixed cost of ownership — which only local inference delivers.
[08]What happens if the internet goes down?
+
What happens if the internet goes down?
Nothing changes. The Copilot, Auto-Reporting Engine and Predictive Sentinel keep working on the LAN. That is the point of running the brain on-prem.