An auditable AI layerover your systems,with the right model for the task.

System topology

Figure 2: the model routing fabric. The incoming task goes into a classifier and router decision node, from there it branches into one of the four model layers (R1 to R4), then verify, a human approval gate and the audit log follow. The distillation loop trains a locally running model from the frontier layer.

• R1

  Local / fine-tuned

  A model running on an on-prem GPU, fine-tuned on your data for high-volume and sensitive work. The cheapest and fastest layer, the data stays in.

• R2

  Frontier reasoning

  For rare cases that need judgment. In the EU region or globally, always through a human gate, at the money and risk points.

• R3

  Specialized code

  A model tuned for code review: it takes every engine change under review and catches the real bugs before release.

• R4

  Vision

  Document and KYC processing: reading scanned documents, IDs and forms into structured data.

• T1Local specialized (chain analytics)

  On-chain AML and sanctioned-address screening

  C5 Settlement orchestration / C6 Custody orchestration

  It screens every deposit and withdrawal before the transaction is finalized: against sanctions lists, known mixers and risky address clusters. On a hit, the withdrawal automatically stops and waits for human review.

  Why: High-volume, repetitive screening where a local model tuned for chain analytics runs cheaply and fast, and the sensitive address data stays in.

• T2Frontier vision and reasoning, human in the loop

  KYC document and PEP adjudication

  C7 Client (onboarding, KYC/AML)

  It reads and cross-checks the onboarding documents, flags the contradictions and PEP or sanctions matches, then prepares a decision proposal. The final approval is always the compliance officer's.

  Why: It needs document understanding and judgment at once, so the frontier layer with visual and reasoning ability, but the decision stays with a human.

• T3Local fine-tuned, human approves

  Settlement and reconciliation discrepancy detection

  C4 Clearing and netting / C5 Settlement orchestration

  It compares expected and actual settlements, surfaces the unmatched item and discrepancy, then proposes a concrete resolution step. The proposal is approved by the clearing team before execution.

  Why: A well-contextualized, repetitive reconciliation pattern that a fine-tuned local model handles accurately and cheaply.

• T4Pattern model and frontier case write-up

  Trade surveillance across venues

  C2 Execution / C3 Position and risk

  From the order flow arriving from multiple venues, it flags suspicious patterns, for example wash trading or spoofing. The frontier layer takes the suspicious cases and prepares them with a reasoned write-up for supervisory review.

  Why: The high-volume screening goes to a fast pattern model, the case write-up to the rarely, more expensively called frontier layer: so cost follows the signal.

• T5Frontier reasoning, a human approves every step

  Treasury and liquidity copilot

  C8 Treasury and liquidity

  It sees the positions and the liquidity situation, proposes reallocation between rails and wallets, and justifies the move. Execution only happens after human approval.

  Why: It needs judgment and seeing the connections, with rare calls, so frontier reasoning, but with a strict human gate because money moves.

• T6Frontier code (a real AI review routine)

  Engine change-risk review on every PR

  Core (platform engine)

  It runs on every change request, reviews the change for risk and regression, and gives concrete comments to the developer. The merge decision is made by the code owner.

  Why: Our live AI code review routine caught real security bugs in open-source projects, which needs the frontier layer tuned for code.

• T7Mid-tier and RAG

  Support and client communication triage

  C7 Client (portal and API)

  It understands and categorizes incoming requests, assembles a draft answer from the internal knowledge base, and escalates the urgent case. The answer is always sent by a human.

  Why: Our internal RAG assistant saves about 100 work hours a week across 24 users, and for triage the cost-effective mid-tier and source-bound RAG are enough.

• T8Agentic frontier, a human executes

  Operations incident copilot

  Ops (observability)

  From the observability data it assembles the picture of the incident, correlates the signals and proposes a concrete intervention plan step by step. The intervention is carried out by the on-call human.

  Why: It calls for multi-step, tool-using reasoning, so an agentic frontier layer, but the actual action is triggered by a human.