Platform
AI does the work, humans release it
xRubber is the AI-native OS for commodity trading. A read-only AI employee reads documents, extracts trades, marks to market and raises warnings; private local models keep your data on-premises; and an event-driven modular core keeps every number decimal-precise, auditable and replayable.
A read-only AI employee
It sits in every channel, answering, summarizing and drafting within each requesting user's permissions. It can propose and it can warn, but a human always reviews and releases — the AI never writes financial data directly, and there is no bypass.
- ✓ Looks up, extracts and warns — the permission to write business data simply does not exist
- ✓ Every answer respects the requesting user's permission scope; out-of-scope data stays invisible
- ✓ A real employee account in the identity system: zero write scope, fully audited, seat-metered
- ✓ Document intake produces a structured draft; a human confirms in-system before anything lands
Read-only tools on one standard protocol
The AI reaches your business through a whitelist of read-only tools — contracts, orders, inventory, exposure, credit lines, margin, market data, reports. They speak standard MCP, and every call carries the requesting user's identity through one policy decision point.
- ✓ Contracts, orders, inventory, exposure, credit, margin, market, reports — all thin read-only wrappers
- ✓ Read versus write is a hard split in code, and the write set is empty
- ✓ Any answer with numbers must come from a tool result with a source link — the model never quotes figures from memory
- ✓ Standard MCP: swap the model, and the tool contracts do not change a line
Memory that learns your desk
The AI remembers your conventions and preferences, not your numbers. Each memory is anchored to a real master-data ID, scoped by permission and time-versioned — and a memory page shows exactly what it learned, from which message, and lets you edit or delete any of it.
- ✓ Remembers conventions, not numbers ("this counterparty prefers CIF"); amounts are always queried live
- ✓ Each memory is anchored to a real master-data ID — no fuzzy guessing about which entity it means
- ✓ Memory carries permission scope: your private thread with the AI stays private
- ✓ When a convention changes, the new one supersedes the old with history kept — the AI never answers with a stale rule
Runs on your own hardware, built for trust
A private local model with RAG keeps every trade, price and counterparty inside your walls — it runs fully offline and starts on a single machine. Underneath, a modular monolith on an event backbone keeps money decimal-precise and every change replayable and auditable.
- ✓ Local LLM plus RAG: private data never leaves the premises, and it runs offline
- ✓ Starts on a single machine with Docker Compose — no cluster, no cloud required
- ✓ Money is always decimal, never float; different currencies never silently sum
- ✓ Modular monolith plus event backbone (Outbox): every change is replayable and audited