Zzeti Zeka PlatformEnterprise AI

AI Agents

Autonomous agents that reason, plan, and act. MCP-powered tool access with built-in guardrails. Detect anomalies, diagnose causes, assign actions, measure impact — closed loop.

AI Gateway

Unified API across 30+ LLM providers. Semantic caching, semantic routing, and token-budget management. Use the right model for the right workload.

Role-Based Access

Hierarchical permission model: Store → Region → Headquarters. Every query, every action, every output — governed by role.

MCP Integrations

Model Context Protocol connectors for the systems you already run: Nebim V3, Google Workspace, Logo, SAP, and more.

Security & Compliance

Enterprise-grade controls: PII masking, WAF, IP firewall, geo-blocking, secret detection. KVKK, GDPR, ISO 27001, SOC 2, HIPAA, PCI DSS aligned.

Workflow Designer

Visual workflow builder. Conditional logic, parallel execution, human-in-the-loop approval. Drag, drop, ship.

Local Models

Run open-weight models on your hardware. Ollama, vLLM, Llama.cpp, LoRA fine-tuning. No token costs for high-volume predictable workloads.

Deployment Flexibility

Run on private cloud, public cloud (AWS, Azure, Huawei), Apple Silicon edge, or commodity NVIDIA GPU hardware. Pick the deployment that matches your sovereignty rules and workload shape.

Runs on NVIDIA DGX Spark

Desktop-class AI inference on NVIDIA's Grace Blackwell GB10 — 1 PFLOP FP4, 128 GB unified memory. Run Zzeti with 70B-parameter open-weight models locally, no cloud token costs, no data leaving your office. Pair two DGX Sparks for 405B-class workloads. Perfect for retail HQs, factories, hospitals, and government — anywhere data sovereignty matters and a rack is overkill.

Context Lake & RAG

A governed data lake for agents — sources, catalog, certified and trusted zones — plus document collections and knowledge bases with routing triggers. Query Bench runs SQL & MDX through MCPs, never a direct database connection.

Human-in-the-Loop

Actions Inbox routes approvals, inputs, reviews, escalations and anomalies to the right person. Agents message people proactively; people approve, correct or stop them — every decision is logged.

AI Judge & Playground

Evaluate agents with an LLM judge against test cases and pass thresholds before they go live. Compare the same agent across different LLMs side by side — streaming, tool calls and cost.

Stop losing sales to empty shelves

Customers don't buy what isn't on the shelf — and the item that sold out in one store is often sitting in another store or in the warehouse.

Zzeti spots stock-outs days ahead at store × product level and drops a replenishment or transfer proposal into the right person's Actions Inbox for approval.

In the system, not on the shelf

Stock that exists on paper but not physically blocks automatic reordering — the item quietly drops out of sales and nobody notices.

Zzeti flags products whose sales flat-line while stock is still on record, assigns a count task to the store and gets the item back on sale once the record is corrected.

Pass cost increases on to price — on time

Reflecting currency and cost increases in prices too late silently erodes margin; raising prices where customers are price-sensitive loses the sale.

Zzeti tracks cost, demand and price together, lists the products whose margin is eroding and those with room to price, and proposes prices together with the reasoning and the query behind them.

Markdowns at the right time and depth

Discount too early and you burn margin; too late and you are left with dead stock at season end. Campaigns nobody measured get repeated.

Zzeti proposes markdown timing and depth from in-season sell-through, measures what past campaigns really returned and flags loss-making ones before they are repeated.

Stop funding products that don't earn

Slow, low-contribution products occupy shelf and warehouse space — space and budget that could go to the products that actually pay.

Zzeti calculates contribution margin and turnover per product and location, and recommends reorder, markdown or delist for each item.

Plan purchasing around cash

Over-buying ties cash up in the warehouse, under-buying loses sales — and buy decisions often rest on last year's numbers and gut feel.

Zzeti combines the demand forecast with payment terms, currency and supplier lead times, and recommends when and how much to buy together with the cash impact.

Reach the right customer before they buy

Sending everyone the same offer burns the campaign budget, and a loyal customer who is drifting away is noticed only after they have gone.

Zzeti predicts who is likely to buy what, and when, from loyalty and CRM data, flags customers who are drifting away and delivers a personal offer over WhatsApp or e-mail.

Set buy and sell prices from market data

Where prices move every day — used vehicles, equipment, commodities — a wrong purchase price kills the profit before the deal starts, and every extra day in stock costs value.

Zzeti reads market and competitor prices alongside your own past deals, proposes a profitable price range for buying and selling, and warns early about stock that is losing value.

Every morning: what changed, why, and who acts

The numbers are in the report, but answering “why did sales drop?” takes days — and by the time a decision is made, the opportunity is gone.

Every morning Zzeti finds the deviations in sales, margin and stock, explains the cause and assigns the action to the right person; management can ask follow-up questions in plain language and see the query behind every answer.

• Expert & worker agents — expert agents for questions and analysis; goal-driven worker agents that plan, execute and report on their own — with supervisors, critique agents and agent-to-agent (A2A) collaboration.
• Skills, goals, prompts & personas — skills teach an agent how to reason about a scenario, goals define what done means, personas change how it answers (executive, analyst, field staff, compliance officer).
• Chat with debug traces — every answer shows the SQL/MDX it ran, function calls, matched skills, RAG sources and tokens used.
• Voice & WhatsApp — talk to agents by voice and deliver them on WhatsApp, not only in the console.
• Messages & Actions Inbox — agents message people and groups proactively; approvals, inputs, reviews, escalations and anomalies land in one inbox.

• Context Lake — a governed lake for agent data: sources, catalog, certified and trusted zones, query and analytics.
• Collections & knowledge (RAG) — document collections and knowledge bases with routing triggers, so the right agent gets the right context.
• Query Bench — SQL & MDX executed through MCPs — never a direct database connection; AI generates and fixes queries.
• Boards — pinned, auto-refreshing dashboards built from agent outputs — KPI cards, charts, risk totals.
• Ready connectors — Nebim V3 (sales, stock, CRM, loyalty, production, data warehouse), Logo, SAP, Microsoft 365, Google Workspace, Jira, Notion, Power BI, MS SQL, MinIO, Qdrant, Trendyol, TCMB EVDS, TÜİK, TEFAS, e-mail, PDF and document export, AWS and Kubernetes.

• LLM registry & AI Gateway — cloud providers and local models behind one API — semantic caching and routing, token budgets, rate limits, API keys.
• Model Hub, local models & GPU servers — pull open-weight models from Hugging Face, deploy with vLLM or Ollama on local, remote or Kubernetes GPUs (including NVIDIA DGX Spark edge nodes), watch VRAM and GPU load.
• Fine-tune Studio — fine-tune deployed models with your own data, then test and run before promoting.
• Load balancers & strategies — several models behind one endpoint — round-robin, weighted or least-latency spreading, failover between models by cost or capability, health checks.
• MCP servers, Inspector & Functions — deploy MCP servers on Kubernetes or connect by URL, inspect their tools, resources and prompts, and write custom Python tools in the built-in Functions IDE.

• Workflows — a visual builder where agents are nodes — start, webhook and schedule triggers, templates, saved graphs.
• Workflow dashboard — schedules, running executions and approvals in one place, with separate human and agent queues.
• Scheduled runs — daily or periodic autonomous scans, checks and reports — without anyone pressing a button.
• A REST API for everything — every agent, worker, workflow, MCP and dashboard is an endpoint — Postman-style API registry, Swagger docs, access keys.
• Webhooks, e-mail & integration proxy — push events into your systems and send notifications; outbound credentials stay in the proxy and egress is controlled.

• Security Hub — user management, an Access Wizard that writes the right IAM policies from four questions, effective-access view, API keys, WAF rules, secrets and a compliance view.
• Guardrails & enforcement — policies on inputs and outputs, PII masking, geo-blocking, IP firewall, secret detection.
• AI Judge, integrity & feedback — LLM-based evaluation with test cases and pass thresholds, integrity checks and a user feedback loop before and after release.
• Monitor — sessions, traces, gateway traffic, function calls, agent fleet, latency by LLM/MCP/agent, success rate, alerts and a live topology map.
• Cost ops — budgets per LLM, monthly limits, rate limiting and cost per request — and zero token cost when work runs on local models.

AI fed by Nebim V3 sales and stock data routes the right product to the right store at the right time. Reduces overstock and stockouts.

Sales forecasts by season, region, and channel. Historical sales data trained into predictive models that lower forecast error.

Automated price and discount optimization based on stock depth, competitor data, and customer behavior. Margin and sales lift together.

Income/expense analysis detects cash flow deviations before they happen. Proactive alerts shorten the cash cycle and reduce logistics costs.

AI fed by Nebim CRM and loyalty data generates customer-specific recommendations and campaigns. Customer lifetime value increases.

Store performance analysis, employee productivity tracking, return/loss analysis, automated reporting. Total operational costs go down.

Quotation in hours, not days

Quotation depends on a few experienced people: reading the drawing and estimating material, machining time and tooling takes days, and slow answers cost orders.

Excel and WhatsApp become structured data

Shift output, scrap and downtime are collected by hand — in spreadsheets or messaging apps — so the numbers arrive late and are hard to trust.

Maintenance that reminds you — not the other way round

Press strokes, furnace hours and mould shot counts are rarely tracked systematically; maintenance is reactive and unplanned downtime arrives at the worst moment.

Visual quality control at the line

Final inspection is manual and visual; defects escape to the customer and come back as PPM figures, claims and rework.

Forecast from the orders you already have

Demand 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.

Know-how that no longer leaves with the foreman

Set-up parameters, customer specifications, quality procedures and the fixes for recurring problems live with a few experienced people and in documents nobody opens.

A management cockpit — without exposing the P&L

Management wants plant KPIs at a glance without exposing financial figures to everyone — so dashboards either show too little or never get built.

Customer drawings never leave the plant

Customer drawings, recipes and prices are confidential — often under NDA — which rules out sending them to public cloud AI services.

Local LLMs on your own hardware — and the platform around the model.

Open-weight models such as Qwen, Llama and Mistral run on your GPU with vLLM or Ollama; Zzeti adds what a company needs around them — role-based access, RAG over your documents, MCP connectors to your ERP, agents, guardrails and an audit trail — with zero external calls if you choose an air-gapped installation. It is the platform for teams in Türkiye that want to work with a local LLM without their data leaving the company.