Eight scenarios for mould, die, press, casting, machining and plastics SMEs — mapped to the Zzeti Zeka Platform, delivered and operated by Skyloop. On your own servers if the drawings must never leave the plant.
In Türkiye's mould, die, press, casting, machining and plastics SMEs the picture is mostly the same: strong production know-how, thin digital foundations. Production is tracked in spreadsheets, orders travel over messaging apps, the ERP is missing or partially used, maintenance is reactive and quotations depend on a few experienced people. These are not AI problems yet — but they are exactly where an agent platform pays off first, because it can read what already exists and put structure around it.
The good news: none of this needs a multi-year digital transformation before AI can help. An agent that reads a spreadsheet, a WhatsApp photo and an ERP table is already useful on day one — and it creates the structured data the next step needs.
Each scenario starts from a situation common in manufacturing SMEs and ends with the Zzeti features that handle it. None of them requires a new ERP, a data-warehouse project or a data-science team.
Quotation depends on a few experienced people: reading the drawing and estimating material, machining time and tooling takes days, and slow answers cost orders.
An agent reads the RFQ e-mail and the drawing, pulls similar past parts and quotes from the ERP or the quote archive, estimates material, cycle time and tooling with your own cost rules and drafts the offer. The engineer reviews and approves in the Actions Inbox; every assumption is visible in the trace.
Worker agent · RAG · Actions InboxShift output, scrap and downtime are collected by hand — in spreadsheets or messaging apps — so the numbers arrive late and are hard to trust.
Agents on WhatsApp take the shift report as a photo, a voice note or a message, extract the fields, validate them against the order and machine list and write them to the Context Lake or the ERP. Boards show OEE, scrap and on-time delivery the next morning — from data that already existed.
WhatsApp agent · Context Lake · BoardsPress strokes, furnace hours and mould shot counts are rarely tracked systematically; maintenance is reactive and unplanned downtime arrives at the worst moment.
Scheduled runs read the counters from the ERP, the PLC export or the shift report, compare them with the maintenance plan and message the responsible person before the threshold — with the checklist attached. Every completed job is logged; the next step is anomaly detection on energy and vibration data.
Scheduled runs · Messages · Anomaly detectionFinal inspection is manual and visual; defects escape to the customer and come back as PPM figures, claims and rework.
A camera at the station and a vision model on an edge GPU — NVIDIA DGX Spark or a standard workstation — flag porosity, cracks, burrs and missing operations in real time. Images and decisions stay on the plant network; the agent files the nonconformity, notifies the shift lead and tracks the trend by mould and machine.
Local models · Edge GPU · WorkflowsDemand planning relies on experience and informal customer signals; raw material is bought late or at the wrong price and capacity swings between idle and overloaded.
Agents combine order history, open orders, customer call-off schedules and market indicators to forecast demand by customer and part family, propose purchase timing and flag capacity conflicts weeks ahead. Planners keep the last word — the agent shows its reasoning.
Forecasting · Query Bench · BoardsSet-up parameters, customer specifications, quality procedures and the fixes for recurring problems live with a few experienced people and in documents nobody opens.
Procedures, manuals, specifications, past nonconformities and set-up sheets become a knowledge base with RAG. Operators ask in Turkish on WhatsApp or a tablet at the machine and get the answer with its source. Role-based access keeps customer documents where they belong.
RAG · Collections · Role-based accessManagement wants plant KPIs at a glance without exposing financial figures to everyone — so dashboards either show too little or never get built.
Boards built from agent outputs show OEE, scrap, on-time delivery, energy per ton, open quotes and overdue receivables in operational terms. Role-based access decides who sees which board and which figure; management asks follow-up questions in natural language and sees the SQL behind every number.
Boards · Role-based access · Query BenchCustomer drawings, recipes and prices are confidential — often under NDA — which rules out sending them to public cloud AI services.
Zzeti runs on your own server, on an NVIDIA DGX Spark or in a private cloud — air-gapped if needed, with open-weight models and zero external calls. Guardrails mask personal data, the integration proxy controls every outbound connection and the audit log shows who asked what.
On-prem · Air-gap · DGX SparkZzeti is the AI layer; Skyloop builds and runs what sits under it. Most plants need a few of these before a pilot can be trusted — and all of them are things a plant can start next week.
3-2-1 backups for the ERP, the file server and the CAD/CAM archive, tested restores and an off-site or cloud copy — so a ransomware incident or a hardware failure does not stop production.
From an ageing on-site server to managed virtualisation, a private cloud or the AWS Istanbul Local Zone; migration and 24/7 operations for Logo, SAP and custom ERPs.
A pipeline that collects PLC, counter, scale and energy data securely into the ERP and the Zzeti Context Lake, with segmentation between the OT network and the office network.
NVIDIA DGX Spark or GPU workstation set-up for visual quality control and local models — inside the plant network, sized for the workload, operated by Skyloop.
VPN or zero-trust access for machine vendors and external service providers, MFA, patch management and endpoint protection with Trend Micro.
Who sees what: access rules for customer drawings, HR data and financial figures, a KVKK inventory and an AI usage policy — paired with the data-governance module of the Executive AI Competency Program.
What lives in the ERP, what lives in spreadsheets, what lives in people's heads — plus backups, network and access. A short scan, one honest picture.
Quotation or maintenance — one measurable process. The agent reads what already exists; nobody replaces a system. Success is a number you agreed on before the pilot.
Skyloop puts backups, ERP integration, edge hardware and data-governance rules in place, so the platform can grow beyond the pilot — on-premise, private cloud or AWS.
Yes — that is the most common starting point in manufacturing SMEs. Agents read spreadsheets, e-mails, PDFs and WhatsApp messages as they are, extract structured data into the Context Lake and validate it. When an ERP exists (Logo, SAP, Nebim or custom), MCP connectors read and write it with approvals; when it does not, the Context Lake becomes the first single source of truth. You do not have to finish an ERP project before starting with AI.
In our experience the fastest returns come from four places: AI-assisted quotation (hours instead of days, less dependence on a few experts), maintenance reminders for presses, furnaces and moulds (fewer unplanned stops), structured shift and scrap data captured from messaging apps and spreadsheets (decisions on numbers people trust) and vision-based quality control at the line (defects caught before the customer). Demand forecasting and a management cockpit follow once the data is flowing.
Yes. Zzeti runs on your own server, on an NVIDIA DGX Spark or in a private cloud, fully air-gapped if required — open-weight models, zero external calls, no telemetry. Customer drawings, recipes and prices never leave the plant network. If you later want cloud LLMs for non-sensitive tasks, the AI Gateway routes only what you allow, with PII masking in front of every model.
Quotation, maintenance, quality or the shift report — bring one. We look at the data you already have and agree on the number a pilot has to move.