HomeAboutProductsR&DBlogGet in Touch
Financial Compliance — Under Development

RLYXA AML

A specialised AML intelligence platform powered by a fine-tuned Qwen2.5-32B-Instruct model. Natural-language transaction analysis, automated SAR drafting, and FATF-cited compliance guidance — deployed entirely within your infrastructure.

Capabilities

What RLYXA AML can do

Unlike generic LLMs, RLYXA AML is trained exclusively on AML/CTF content. It cites specific FATF Recommendations, explains its reasoning, and runs locally on bank infrastructure.

Analyzer Chat

Type natural language queries like "Check transaction £9,500 to Country X" and receive a structured compliance assessment with FATF citations.

Spider Mode

Upload CSV transaction files for batch analysis. The system generates a network graph of relationships, flags suspicious patterns, and produces a consolidated risk report.

SAR Drafting

Auto-generates Suspicious Activity Report narratives with proper structure, regulatory context, and supporting evidence references.

FATF Flags

Every output references specific FATF Recommendations. The model was trained on the full FATF recommendation corpus and explains which rule applies to each scenario.

PEP & KYC Guidance

Answers enhanced due diligence questions for Politically Exposed Persons, sanctions screening, and high-risk jurisdiction onboarding.

QA Queue

Built-in FCA-style 5% human sampling queue. Low-risk transactions are flagged for analyst confirmation, ensuring regulatory compliance without reviewing every case.

Model Architecture

Private, explainable, specialised

RLYXA AML is built on Qwen2.5-32B-Instruct with a 33.5 MB LoRA adapter — fine-tuned on 11,863 specialised instruction-output pairs covering FATF recommendations, transaction analysis, SAR drafting, and sanctions compliance.

Fine-Tuned, Not Generic

Trained exclusively on AML/CTF content. It does not hallucinate general knowledge — it cites FATF Recommendations, explains step-by-step, and identifies legitimate "No SAR required" cases.

Runs Locally

Deployed entirely within your data centre or VPC. No data leaves your network. Optional air-gapped deployment with no internet connectivity required.

Explainable by Design

Every response includes the specific FATF Recommendation cited, the reasoning chain, and the confidence level. Auditors get transparency, not black boxes.

Fast & Cost-Effective

Inference in 3–5 seconds per query. The adapter is only 33.5 MB on top of the open-source base model — no per-token API costs, no vendor lock-in.

Technical Specifications

Base ModelQwen2.5-32B-Instruct
Fine-Tuning MethodLoRA (4-bit QLoRA)
Training Data11,863 instruction-output pairs
Adapter Size33.5 MB
Context Length128K tokens (max)
Inference Speed~3–5 seconds per query
GPU Requirement24 GB VRAM (min) / 48 GB+ (rec)
DeploymentOn-premise or VPC (air-gapped OK)
Example Queries

What you can ask

What does FATF Rec 19 require for high-risk jurisdictions?

Customer deposits $9,800 daily for 5 days. Is this suspicious?

A former minister wants to open an account. What EDD steps?

Write a SAR narrative for a structuring case

What are red flags for crypto money laundering?

What is the OFAC 50% Rule?

Roadmap

Where we are heading

v1.0 (Current)

Core AML Q&A, FATF citations, SAR drafting, Spider mode batch analysis

v1.5

RAG system with FATF PDF ingestion, enhanced citation accuracy

v2.0

Multi-language support (Arabic, French, Spanish)

v2.5

API endpoints, real-time batch transaction analysis

v3.0

Autonomous AML agent with tool calling and proactive alerting

Interested in RLYXA AML?

We are actively developing RLYXA AML and welcome early design partners from financial institutions, compliance teams, and regulatory technology providers.