One on-premise platform. Multi-agent industrial AI.
Sakaki packages a local LLM, a multi-agent orchestration layer, an enterprise knowledge base and a Text-to-SQL engine into one appliance — pre-integrated with SQL Server, PostgreSQL, MySQL and Oracle, delivered air-gapped and ready for the shop floor.
[01]Module 01
NVIDIA Performance Tier
Built on NVIDIA's mature CUDA ecosystem — ideal for enterprises that value development speed, model versatility, and the option to expand into many different AI tasks over time.
For scenarios that demand fast generation and deep logical reasoning, run mainstream open-source LLMs (Llama 3 / Qwen 2.5) directly for complex industry-report analysis and long-context retrieval (RAG). Sub-second SQL synthesis and multi-step diagnostics, with no cloud calls. RAPIDS acceleration library speeds up Pandas data cleaning and Scikit-learn machine learning model computation, enabling minute-level processing of billions of data points.
Key selling points
▸Run 8B–70B local LLMs for complex reasoning and long-context reports
▸Sub-second SQL synthesis and multi-step diagnostics on the LAN
▸Rack or desktop form factor to fit machine rooms and engineering offices
▸No cloud calls, no token burn, full data sovereignty
Typical scenarios
▸High-mix assembly lines that need instant answers across many SKUs
▸Engineering root-cause analysis over years of MES and alarm history
▸Plants where rich reasoning and low latency must coexist on-prem
[02]Module 02
Edge ASIC Efficiency Tier
Multiple cost-effective edge chips that rely entirely on algorithmic techniques — model quantization, aggressive pruning, and on-device RAG — to deliver domain-specific capabilities. Purpose-built for cost-sensitive SMEs and integrators.
How it's done: aggressive model quantization, tiny expert models (MoE / Tiny Models), lightweight vector retrieval (Lightweight RAG), and distributed task partitioning.
Key selling points
▸Quantized 1.5B–3B Text-to-SQL specialists on dedicated NPUs
▸Fanless, low-wattage design for dusty shop-floor cabinets
▸Same software stack and API as the NVIDIA tier
▸Priced for SMEs and integrator-bundled rollouts
Typical scenarios
▸Cost-sensitive SMEs that want AI without cloud subscription fees
▸Integrators who need a compact, reliable edge node to bundle with machines
▸Production lines where fanless operation and low heat matter
[03]Module 03
Enterprise Knowledge Base (Agentic RAG)
Automatically parses complex internal documents — equipment manuals, operating procedures and more — into a searchable knowledge space.
When a fault occurs at a workstation, vector retrieval delivers the standard repair steps and mistake-proofing instructions in seconds — answers stay consistent with your own manuals and engineering documents.
Key selling points
▸Auto-parses SOPs, EU Machinery Directive and PLC ladder manuals
Typical scenarios
▸Line-side fault: operator gets the exact repair procedure instantly
▸New engineer onboarding against decades of tribal documentation
▸Compliance lookups against EU Machinery Directive clauses
[04]Module 04
Natural-Language Data Copilot (LLM + Text-to-SQL)
Ask questions in plain language against SQL Server, PostgreSQL, MySQL or Oracle — the SQL Agent generates the query automatically and the LLM explains the results.
The SQL Agent turns business intent into JOIN-correct, scope-constrained optimized SQL; RAG fusion over your knowledge base grounds every figure with its source and delivers a full analytical result — not just a table.
Key selling points
▸Text-to-SQL across SQL Server, PostgreSQL, MySQL and Oracle
▸Agentic RAG grounds answers in SOPs, manuals and historical reports
▸Every answer includes the SQL and sources it used
▸Multilingual interface: Swedish, English and Chinese
Typical scenarios
▸Operators asking how to calibrate a station during changeover
▸Shift handover questions that used to chase tribal knowledge
▸Audit traceability: who changed what, when, and which orders were affected
[05]Module 05
AI Fault Report Generation
Automatically produce fault summary, root cause, downtime and KPI reports — exported to PDF, Word or Excel and delivered by email.
The Report Agent assembles fault summary, root-cause analysis, statistical trends, downtime analysis, maintenance recommendations, preventive actions and a KPI dashboard into a single professional report. Trigger on schedule, on alarm event, or on demand; delivered to the responsible person by email or push notification.
Key selling points
▸Auto root-cause analysis with traceable SQL for every figure
▸One-click export to PDF, Word or Excel
▸Email or push delivery to the right people
▸Triggered by schedule, alarm event or natural-language request
Typical scenarios
▸Unattended night shifts: morning briefings ready before the stand-up
▸Customer audits: audit-ready documentation in minutes, not days
▸Changeover validation: catch bad changeovers before scrap accumulates
[06]Module 06
Predictive Analytics & PdM (Deep Learning)
Mines historical fault data with site-tuned deep-learning models to identify recurring failures, equipment degradation and quality risks, and issue early warnings.
Mines historical fault data with site-tuned deep-learning models to identify recurring failures, equipment degradation and quality risks, and issue early warnings.
Key selling points
▸Site-tuned deep-learning models trained on your own history
▸Per-station Risk Score with minutes-to-hours lead time
▸Combines MES history with edge signals like cycle time drift and retries
▸Turns emergency stops into planned maintenance windows
Typical scenarios
▸Tightening-gun wear: schedule replacement at the next changeover
▸Scanner degradation: fix the reader before the cascade of timeouts
▸Upstream logistics: refill kanbans before material starvation stops the line
[07]Module 07
Air-gapped Sandbox
Hardware-level isolation: the AI never touches the live production database — zero write-back risk, zero data egress.
An on-device sync service mirrors the factory SQL Server into an internal sandbox physically isolated from the production environment in near real-time. All agent reads happen inside the sandbox — no data ever leaves the box.
Isolation
Physical / on-device
Sync
Near real-time
Cloud egress
None
Key selling points
▸Physical isolation: AI reads a sandbox mirror, never the live MES
▸Near real-time sync keeps answers fresh without loading production
▸Zero write-back risk and zero cloud egress
▸Works fully offline once deployed
Typical scenarios
▸Security-critical plants with strict no-cloud policies
▸GDPR-sensitive lines where data must never leave the site
▸Air-gapped networks where WAN access is limited or forbidden
[08]Module 08
Custom AI Applications
Whether it's visual quality inspection or a deeply integrated dedicated Agent, we deliver production-ready private AI applications tailored to your on-site workflows, data shape and compliance requirements.
Whether it's visual quality inspection or a deeply integrated dedicated Agent, we deliver production-ready private AI applications tailored to your on-site workflows, data shape and compliance requirements.
▸Fully private deployment; models and data stay inside your network
▸Seamless integration with Sakaki appliance or existing industrial IT
▸Supports vision, time-series, text, and voice use cases
Typical scenarios
▸Unique quality standards: custom vision-defect model embedded in the line
▸Private MES/ERP integration: bespoke data Agent and report pipeline
▸Strict compliance: custom architecture aligned with your air-gap and data policies
// Use cases
Scenarios & Efficiency
[01]
Conversational Data Copilot
Case 01
New-operator onboarding
New operators skip paper manuals and learn by interacting with the AI Enterprise Knowledge Base.
Expected benefit
Cuts onboarding time from weeks to days; reduces first-month quality deviations by removing the guesswork from tribal knowledge.
Case 02
Engineering-change traceability
After an ECN, the line lead asks why Station 02 cycle time shifted. The Copilot pulls the change document, correlates it with cycle_log data, and explains which parameter changed and which work orders are affected.
Expected benefit
Turns 'who changed what' questions from hours of ticket chasing into a 30-second conversation.
Case 03
Shift-to-shift knowledge retention
Night-shift fixes for recurring scanner misreads are automatically summarized into the knowledge base. Day shift asks about the same symptom and gets the proven workaround instantly.
Expected benefit
Stops the same problem from being solved twice; preserves retiring operators' expertise in a searchable, queryable form.
[02]
Auto-Reporting Engine
Case 01
Unattended night shifts
At 02:14, if the alarm rate spikes, the appliance automatically generates a brief and sends it via email or SMS to the duty manager.
Expected benefit
Replaces morning-after firefighting with pre-shift action items; reduces mean time to understand (MTTU) by 60–80%.
Case 02
Customer audit preparation
Ahead of a quality audit, the plant manager requests a weekly summary of top stoppages, corrective actions and repeat occurrences. The engine generates a markdown/PDF report with traceable SQL for every figure.
Expected benefit
Audit-ready documentation in minutes instead of days of manual Excel consolidation.
Case 03
Changeover validation
After a product changeover, first-hour alarm rate is 3× the historical baseline for the same SKU. The engine produces a before/after comparison report flagging which stations deviated most.
Expected benefit
Catches bad changeovers while they are still cheap to fix, before scrap and downtime accumulate.
[03]
Predictive Sentinel
Case 01
Tightening-gun predictive swap
Risk Score for the tightening tool climbs from 22 to 71 over 90 minutes. Cycle time drift and retry-rate anomalies point to hardware degradation. The Sentinel warns the maintenance team with a 6-hour lead time.
Expected benefit
Avoids unplanned stoppages; lets maintenance schedule the swap during the next planned changeover instead of an emergency line stop.
Case 02
Scanner degradation watch
the scanner re-read rate is 3.2× baseline but has not yet triggered any alarm. The Sentinel flags the station as scanner-degrade risk before the first timeout occurs.
Expected benefit
Prevents the re-read storm from starving the station of part IDs and triggering a cascade of cycle timeouts.
Case 03
Upstream logistics bottleneck
cycle time micro-drift and heartbeat jitter upstream of Station 03 indicate a kanban refill delay. The Sentinel estimates 25–60 minutes until material starvation.
Expected benefit
Gives logistics time to refill before the line stops; turns material alarms from surprises into scheduled interventions.
// Next
Pilot Sakaki on your line.
We are onboarding a small cohort of SME factories and MES integrators for the first deployments. Tell us about your environment.