HUMAN-CENTERED AI FOR FINANCIAL FORENSICS
Human-Centered LLM-Agent System for Detecting Anomalous Digital Asset Transactions
We present HCLA, a human-centered multi-agent system for anomaly detection in digital-asset transactions. The system links three roles—Parsing, Detection, and Explanation—into a conversational workflow that lets non-experts ask questions in natural language, inspect structured analytics, and obtain context-aware rationales. Implemented with an open-source web UI, HCLA translates user intents into a schema for a classical detector (XGBoost in our prototype) and returns narrative explanations grounded in the underlying features.
HCLA: Empowering Financial Integrity with Interpretable AI
Leveraging advanced AI to bring clarity and trust to digital asset anomaly detection, HCLA provides robust performance with unparalleled transparency.
0
Accuracy (XGBoost Baseline)
0
Precision
0
F1-Score
0
Avg. Response Latency
↑
Interpretability Enhanced
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
HCLA Framework Overview: Conversational Anomaly Detection
The Human-Centered LLM-Agent (HCLA) framework operationalizes anomaly detection as a conversational workflow between human users and AI agents. It integrates a Parsing Agent (ChatGPT), a Detection Agent (XGBoost), and an Explanation Agent (Gemini) into a seamless, auditable pipeline. This modular design ensures transparency and adaptability.
Enterprise Process Flow: HCLA Data Transformation Pipeline
User Input (Unstructured Text)
→
Structured Query (JSON Format)
→
Raw Analysis Result (Numerical Score)
→
Model Result Interpretation & Synthesis
The Parsing Agent converts natural-language queries into structured JSON schemas. The Detection Agent then uses these schemas to compute anomaly probabilities based on temporal, transactional, and graph-connectivity features. Finally, the Explanation Agent translates numerical scores into plain-language reasoning, supporting iterative questioning and human understanding.
Detector Performance & User Validation
The HCLA framework maintains strong anomaly detection accuracy using an XGBoost baseline, while significantly enhancing interpretability and user trust, validated through a simulated user study.
0
Accuracy
0
Precision
0
Recall
0
F1-score
p < .001
Paired t-test results showed HCLA explanations significantly higher in Trust and Clarity.
In a simulated user study with a Micro-Expert Panel, participants rated HCLA explanations significantly higher in both trust and clarity (p < .001) compared to purely numerical outputs from XGBoost. This empirically validates the effectiveness of HCLA's AI explanation model.
Human-Centered Design Principles
HCLA demonstrates several key advantages derived from its human-centered AI design, fostering transparency, adaptability, and inclusivity in complex analytical workflows.
Key HCLA Advantages
Accessibility: Natural-language interaction removes technical barriers, enabling non-experts to analyze transactions without coding or schema creation. Queries like "Analyze transactions from my wallet over the past week" are processed autonomously.
Interpretability: Context-rich narrative explanations translate complex model outputs into understandable reasons, such as "repeated high-value transfers to unverified counterparties during off-peak hours," fostering user confidence.
Trust and Transparency: The multi-agent structure decomposes reasoning steps, making each stage visible and queryable. Users can audit the pipeline, building trust essential for regulatory and compliance contexts.
Identified Limitations
Despite its strengths, HCLA faces challenges:
- Computational Cost and Latency: LLM usage introduces overhead, limiting real-time monitoring for high-frequency streams.
- Domain-Specific Adaptation: Generic LLMs can sometimes produce ambiguous interpretations; fine-tuning with blockchain and finance corpora is needed.
- Scalability Constraints: Scaling to continuous blockchain streams will require asynchronous orchestration and caching.
Benchmarking Against Existing Solutions
HCLA distinguishes itself from prior LLM-based anomaly detection systems by offering sustained dialogue, modularity, and reproducibility under prompt contracts, addressing key limitations in one-way reporting or single-agent approaches.
| Study | Summary | HCLA's Advantage |
|---|---|---|
| RAAD-LLM (Russell-Gilbert 2025) | LLM retrieves relevant context via RAG, summarizes results. | Handles input at "event description" level, lacks sustained conversational interaction. HCLA enables full conversational interaction. |
| LLM-Augmented Explanations (Watson 2024) | LLM reformulates graph model outputs into human-readable explanations. | One-way reporting; users cannot interactively explore/refine reasoning. HCLA offers iterative refinement. |
| CALM (Devireddy & Huang 2025) | LLM judges time-series anomaly detectors, offers adaptive summaries. | System-centric design focused on model performance, not user interaction. HCLA prioritizes user interpretability. |
| AnoLLM (Tsai et al. 2025) | Tabular records serialized into prompts, analyzed by LLMs for descriptive outputs. | Provides explanations but lacks conversational interface and modular structure. HCLA offers full modularity (Parsing-Detection-Explanation). |
| Anomaly Detection for Short Texts | Users paste logs/short texts; LLM summarizes cause and detection results. | Single-agent pipeline; no explicit modular separation. HCLA uses multi-agent system for extensibility and control. |
| HCLA (proposed framework) | Integrates Parsing, Detection, and Explainer agents through Gradio, stabilized by prompt engineering. | Enables conversational interaction, modular reproducibility, and transparent reasoning for non-expert users. |
Calculate Your Potential AI Impact
Estimate the annual savings and efficiency gains your organization could achieve by implementing a human-centered AI system like HCLA.
Estimated Annual Savings
$0
Hours Reclaimed Annually
0
Your Path to Transparent AI for Financial Forensics
Implementing HCLA for digital asset transaction monitoring involves a structured approach to ensure seamless integration and maximum impact.
Phase 1: Discovery & Strategy
Initial consultation to understand current anomaly detection workflows, data sources, and compliance requirements. Define key objectives and scope for HCLA integration.
Phase 2: Data Integration & Model Adaptation
Securely integrate digital asset transaction data. Adapt and fine-tune LLM agents and the detection model (e.g., XGBoost, GNNs) to your specific dataset and domain terminology.
Phase 3: Customization & UI Deployment
Customize narrative explanation styles, implement specific user interaction loops, and deploy the Gradio-based web interface within your secure environment.
Phase 4: Training & Operationalization
Provide training for financial analysts and compliance officers. Establish monitoring protocols and continuous improvement feedback loops for agent refinement.
Phase 5: Advanced Integration & Scaling
Explore real-time monitoring extensions (Kafka/Flink), multimodal data fusion (logs, screenshots), and large-scale user validation for enterprise-wide deployment.
Ready to Transform Your Anomaly Detection?
Don't let opaque systems hinder your financial forensics. Partner with us to implement HCLA and bring unprecedented transparency and trust to your digital asset operations.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat