Amoza AI Platform — one platform serving every pillar
Built for aPlanner first. Designed for MAKE, MOVE, SELL and DELIGHT.
Five components — aCore, aFlow, aConnect, aExecutor, aData — that every Amoza product is built on. Today they power one product, aPlanner (pre-commercial MVP); the plan is to productize the platform in 2028–29 so that each product added to the chain starts with the data and AI of the one before.
One data model. Every pillar.
Production → inventory → orders → customer signals. One shared data model and one AI layer, so every product added to the chain starts with the data and AI of the previous one — from shop floor to customer door.
Five components and their role for every pillar
Status labels are honest: "In development" means the component exists inside the aPlanner MVP; "Roadmap" means design only.
aCore
Role for every pillar: Forecasting for the fab today — for inventory, pricing and churn tomorrow.
The model layer: ensemble forecasting (LSTM, Prophet, XGBoost, Transformer), constraint-based optimization (OR-Tools / MILP) and Monte-Carlo simulation. Today aCore is the engine inside aPlanner; the same training and inference pipelines are designed to serve every later product.
Python, PyTorch, XGBoost, OR-Tools / MILP
aFlow
Role for every pillar: One ingestion layer for MES, ERP, WMS, commerce and CX data.
Ingests, validates and transforms data from the systems a fab already has — MES, SAP ERP, yield / bin databases and planner spreadsheets — into one planning data model. Built first for aPlanner; designed to carry inventory, order and customer data later.
Python, PostgreSQL + TimescaleDB
aConnect
Role for every pillar: Intelligent data mapping between enterprise systems and the Amoza data model.
Pre-built connectors — SAP, Oracle, in-house MES — in development, read-only first. Write-back of approved plans only after planner sign-off. Commerce and CRM connectors follow when the SELL and DELIGHT pillars open.
FastAPI, Kubernetes on-prem or AWS Seoul
aExecutor
Role for every pillar: Multi-step workflows across pillars: re-plan → re-allocate → notify customer.
Agents that run multi-step workflows across pillars once more than one product is live — for example re-planning production, re-allocating inventory and notifying the customer in one flow. Always with human-in-the-loop approval. Roadmap item; not part of the aPlanner MVP.
aData
Role for every pillar: Required for Korean fabs and for customer data alike.
Governance for the shared data model: master data, lineage, quality scoring and PII detection. Supports the K-ISMS / ISO 27001 roadmap and the data-residency requirements of Korean fabs. Roadmap item; the MVP ships the minimum governance aPlanner needs.
Shared AI capabilities — built once in aPlanner, reused later
All figures are design targets from simulation on synthetic fab data; no customer deployments yet.
Demand Forecasting
Ensemble models; target +20–30 pts forecast accuracy vs. spreadsheet planning
Constraint Optimization
Yield-aware, constraint-based scheduling; target full-fab schedule in < 30 min
Scenario Simulation
Monte-Carlo what-if engine; target minutes per scenario
Anomaly Detection
Detection of production, logistics and behaviour anomalies
Predictive Modeling
Equipment health, delivery ETA, churn prediction
NLP / Language AI
Natural-language search, sentiment analysis, agent assist
Recommendation
Product, content and action recommendations
Autonomous Agents
Multi-step workflow execution across pillars
Build on the platform (roadmap)
Developer access follows platform productization (2028–29). Until then, integration is delivered with the founder and Korean SI / MES partners during the pilot.
Step one is aPlanner. Built in Korea.
The platform is the long-term direction. Today we build and validate its first product with Korean fabs and suppliers in the 2027 pilot program.