Built in layers. Designed to work together.
The POC breaks AI-native order management into testable technical components.
MEDVION
Fine-Tuning with LoRA
The POC adapts Llama 3 8B to OMS-specific examples using small trainable layers.
RAG with pgvector
Historical cases are embedded and retrieved for context-aware decisions.
Five Guardrails
Deterministic checks protect execution.
Order Agent
Connects the model, RAG and guardrails.
Feedback Loop
Human corrections can become future training examples.
Service Layer
The POC guide includes FastAPI and Uvicorn in the environment stack.
OMS use cases
Designed around operational decisions.
The training plan covers routing, payment decisions, exception handling, returns/refunds, business configuration and general OMS knowledge.
Routing
Warehouse selection, split/consolidate, priority and backorders.
Payments
Gateway routing, retries, fraud flags and refunds.
Exceptions
Stock, payment, address and carrier issues.
Returns
Eligibility, refunds, restocking and exchanges.