ARIA-v1 — Agentic Risk Intelligence for AML

ARIA-v1 is a fine-tuned Gemma 4 4B model purpose-built for AML (Anti-Money Laundering) compliance analytics. It acts as an agentic tool-calling assistant that routes user queries to the correct analytics function, calls it with correctly structured parameters, and interprets the pre-computed results without hallucinating numbers.

Fine-tuning on a bank's own alert history continuously improves ARIA's precision, reducing the false positive burden on AML investigators over time.

Model Details

Property	Value
Base model	google/gemma-3-4b-it (via unsloth/gemma-4-E4B-it)
Fine-tune method	QLoRA (LoRA r=16, α=32) with Unsloth SFT
Training examples	933 domain-specific AML analytics conversations
Quantization	Q8_0 GGUF
Context length	8,192 tokens
Inference	Ollama (OpenAI-compatible API)
License	Modified OpenRAIL-M — see LICENSE_MODEL

What it does

The model learns four capabilities from the training data:

1. Intent classification — determines which analytics agent handles the query (threshold tuning, segmentation, policy Q&A, OFAC screening, or out-of-scope)
2. Tool calling — emits correctly structured JSON tool calls for Python analytics functions
3. Result narration — copies pre-computed numbers verbatim into the response (eliminates hallucinated figures — a key requirement in compliance contexts)
4. Policy Q&A — answers AML regulatory questions from retrieved document chunks without fabricating citations or statute numbers

Agents and Tools

Agent	Tools
ThresholdAgent	threshold_tuning, sar_backtest, rule_2d_sweep, list_rules, rule_sar_backtest, cluster_rule_summary
SegmentationAgent	ds_cluster_analysis, cluster_analysis, alerts_distribution
PolicyAgent	ChromaDB RAG over FFIEC, Wolfsberg, FinCEN, FATF, EU AMLD 4/5/6, AMLR 2024
OFACAgent	ofac_screening, ofac_name_lookup

How to Use

With Ollama

# 1. Download the GGUF and create a Modelfile
cat > /tmp/Modelfile.aria << 'EOF'
FROM aria-v1-q8.gguf

PARAMETER num_ctx 8192
PARAMETER temperature 0.1
PARAMETER top_p 0.9
PARAMETER stop <turn|>
PARAMETER stop <eos>

TEMPLATE """
{{ if .System }}<|turn>user
{{ .System }}<turn|>
{{ end }}
{{ range .Messages }}
{{ if eq .Role "user" }}<|turn>user
{{ .Content }}<turn|>
<|turn>model
{{ else if eq .Role "assistant" }}{{ .Content }}<turn|>
{{ else if eq .Role "tool" }}<|turn>tool
{{ .Content }}<turn|>
<|turn>model
{{ end }}
{{ end }}
"""
EOF

ollama create aria-v1 -f /tmp/Modelfile.aria
ollama run aria-v1

With the full agentic app

Clone the agentic-aml repository and follow the setup instructions in the README. The app provides a Plotly Dash UI that wires the model to the analytics tools.

git clone https://github.com/speri-psg/agentic-aml
cd agentic-aml
pip install -r requirements.txt
export OLLAMA_MODEL=aria-v1
python application.py

Training Details

• Base model: unsloth/gemma-4-E4B-it
• Method: QLoRA with Unsloth — LoRA r=16, α=32, dropout=0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Epochs: 3 | Effective batch size: 8 (batch=2, grad_accum=4)
• Learning rate: 2e-4 with cosine scheduler, 5% warmup
• Hardware: NVIDIA RTX 3090 / RTX 4090 / RTX 5090
• Inference requirements: 16 GB VRAM minimum (model is 8 GB+ at Q8_0; 16 GB leaves headroom for context and KV cache)
• Training data: 933 examples across threshold tuning, segmentation, policy Q&A, OFAC screening, greetings, and out-of-scope handling

Data Privacy and Air-Gap Deployment

ARIA-v1 is designed to run entirely on-premises inside a bank's own infrastructure. All analytics — threshold sweeps, segmentation, SAR backtests, OFAC screening — are executed by local Python functions against the bank's own data. The model never sends customer data, transaction records, or alert information to any external API or cloud LLM.

When a bank deploys ARIA-v1 on their own hardware, customer data stays within their environment at all times. There is no dependency on OpenAI, Anthropic, or any third-party inference service. This makes ARIA-v1 suitable for regulated institutions where data residency, confidentiality, and audit requirements prohibit sending customer data outside the bank's perimeter.

Intended Use

ARIA-v1 is designed for AML compliance teams and researchers building or evaluating AI-assisted transaction monitoring tools. It is intended to be used with the agentic-aml application stack which supplies the analytics tools and data layer.

Limitations

• Designed for the synthetic dataset schema provided in agentic-aml — adapting to a different schema requires re-mapping column names via column_map.yaml
• Policy Q&A quality depends on the documents ingested into ChromaDB — only documents in the knowledge base can be cited
• Not a substitute for qualified AML compliance advice

License

This model is released under a modified OpenRAIL-M license. Free for personal use, academic research, and organizations with annual revenue and total funding each below USD $2 million. Commercial use above that threshold requires a separate license. The model may not be used to build a competing AML transaction monitoring or financial crime analytics product or service.

Full license: LICENSE_MODEL

Contact

For commercial licensing, enterprise enquiries, demos, or deep dives, open an issue on GitHub or email speri@princetonstrategygroup.com.