CATAS (Compliance And Treasurer Agentic Solutions)

A Multi-Agent System Design for Compliance & Treasury Automation built for the Lyzr Architect Hackathon.

Overview

CATAS uses a Python-orchestrated Dual-Agent framework to solve the reconciliation bottleneck, the audit-trail gap, and the treasury-compliance silo.

The system utilizes deterministic offline ML models to ground LLM reasoning across two distinct operations:

• Agent 1: Treasury Operations (Isolation Forest Anomaly Detection & Prophet Time-Series Forecasts)
• Agent 2: Compliance Operations (Logistic Regression EU Sanction filtering & Immutable Audit Log generation)

Prerequisites

• Python 3.10+
• Lyzr Architect account / access
• Dependencies from requirements.txt

Collaborator Setup Instructions

1. Clone the repository:

   git clone <repository_url>
   cd CATAS

2. Set up a virtual environment:

   python3 -m venv venv
   source venv/bin/activate  # On Windows use `venv\Scripts\activate`

3. Install dependencies:

   pip install -r orchestration/requirements.txt

Pre-Hackathon Pipeline (Data & Offline ML)

To ensure reliable performance during the hackathon, we are using synthesized data and offline-trained ML models. These must be prepared before the live event.

1. Generate Data: Run the data generation script to build synthetic historical data, GL balances, and the live demo dataset.

   # Synthetic data already pre-loaded into /data/ directory

2. Train Models: Train and serialize the Machine Learning .pkl models based on the offline synthetic data.

   # Offline Models already serialized in /models/ directory

Live Execution Pipeline (Hackathon Demo)

CATAS runs seamlessly using a custom Python script that strings together the Offline ML Models with the live Lyzr Agent API.

1. Set your API Keys: Create a .env file in the root directory:

   LYZR_API_KEY="your-api-key"
   LYZR_TREASURY_AGENT_ID="treasury-id"
   LYZR_COMPLIANCE_AGENT_ID="compliance-id"

   • LYZR_API_KEY — from Lyzr Studio → Settings → API Keys.
   • LYZR_TREASURY_AGENT_ID / LYZR_COMPLIANCE_AGENT_ID — the IDs of the two agents you create in Lyzr Studio. Each ID appears in the agent's URL (studio.lyzr.ai/agent/<ID>).

   The exact agent configuration (name, model, role, goal, instructions) is documented step-by-step in docs/CATAS_Architect_Setup_Guide.md §2. Without these three variables set, run_catas.py automatically falls back to mock mode. See orchestration/.env.example for the full list of supported variables (webhook alerts, AWS SES, path overrides).

2. Run the Orchestrator: To run the script in deterministic Mock mode (no Lyzr Credits required):

   python orchestration/run_catas.py --mode mock

   To run the script against the live Lyzr API:

   python orchestration/run_catas.py --mode live --limit 5

3. View the Dashboard: To visualize the generated /logs/audit_trail.txt, load it into the Architect Beta (architect.new) UI dashboard!

Hackathon Phases & Roles

• Data Scientist/ML Engineer: Responsible for features, data schema, and predictive analytics.
• Platform/Orchestration Engineer: Responsible for Python integration, Lyzr API orchestration, AWS Guardrails, and Architect.new UI display.

See docs/CATAS_Hackathon_Roadmap.md for specific timelines and event details.