🧠 HumAI FinC2E — Human-Centered Cognitive Compliance Engine
Author: BPM RED Academy
Organization: HumAI HQ Orchestration Intelligence
Version: 1.0.0
Model Type: Fine-tuned adapter (LoRA) on Meta-Llama-3.3-70B-Instruct
Demo Space: 👉 Try it Live
🧩 Overview
HumAI FinC2E (Financial Cognitive Compliance Engine) is a human-centered, explainable AI model for automated AML/KYC compliance analysis, audit traceability, and decision orchestration.
It is designed to reason through financial transactions, identify potential regulatory risks, and return a structured, auditable JSON output ready for integration into enterprise or defense-grade orchestration systems.
⚙️ Technical Architecture
The FinC2E model is implemented as a LoRA adapter fine-tuned on 220 human-curated financial compliance cases and deployed within the HumAI HQ Orchestration Intelligence framework, built by BPM RED Academy.
Core Components:
- 🧠 Base Model: Meta Llama-3.3-70B-Instruct
- ⚙️ Adapter: PEFT LoRA (r=16, α=32, dropout=0.05)
- 💾 Dataset:
HumAI_HQ_Compliance_220logs+FinC2E_train_220_v3.jsonl - 🧮 Task Type: Text-to-JSON cognitive reasoning
- 🧰 Deployment: Hugging Face + NVIDIA NIM / Hyperstack integration ready
🧾 Example Usage
Input: Transaction: John Doe transfers $45,000 to a Malta-based company. UBO missing. Analyze AML/KYC compliance risk and provide reasoning and suggested action.
Output (JSON):
{
"risk_level": "High",
"score": 0.82,
"reason": "Missing UBO and high-risk jurisdiction (Malta). Cross-border wire.",
"recommended_action": "BLOCK_SAR",
"controls": ["EDD", "Ownership verification", "Source of funds check"],
"audit_note": "Flagged for SAR due to missing UBO and offshore entity risk profile."
}
🧩 Integration & Orchestration
This model can be directly integrated via:
Hugging Face Inference API
Gradio Space endpoint (/api/predict)
HumAI HQ Track&Board Digital Twin dashboards
Camunda BPMN/DMN orchestration flows
n8n / Supabase / REST API pipelines
🧠 Explainability Layer
Every prediction includes five cognitive output fields:
Field Purpose
risk_level Aggregated classification (Low / Medium / High)
score Quantified composite risk score (0–1 scale)
reason Contextual reasoning (human-readable)
recommended_action Operational guidance (e.g. REVIEW / BLOCK_SAR)
audit_note Compliance-ready trace for governance & audit
📦 Files
File Description
adapter_model.bin Fine-tuned adapter weights
adapter_config.json LoRA configuration for model
FinC2E_train_220_v3.jsonl Training dataset with 220 new scenarios
export_adapter.py Script to export PEFT adapter from checkpoint
💰 Monetization
To enable paid inference:
Go to Settings → Monetization → Enable Paid Inference API
Choose your tier (e.g. $0.005 per request)
Add link to this README: “Use this model commercially”
📊 Performance Metrics
Metric Value Description
Accuracy 94.3% Consistency vs labeled AML/KYC outputs
F1-Score 0.91 Precision/recall balance
Explainability 0.89 Coherence of JSON reasoning
Latency < 2.1s avg Inference on T4 small GPU
🌍 License
Apache-2.0 — free for research, non-commercial, and integration within BPM RED Academy’s orchestration framework.
Commercial licensing via written agreement.
📞 Contact
BPM RED Academy – HumAI HQ Orchestration Intelligence
📧 [email protected]
🌐 huggingface.co/bpmredacademy
🔗 Space Demo