eddyzzl/marvis-risk-agent

What it solves

MARVIS-Agent is a local-first platform designed for credit-risk analysis and development. It replaces fragmented workflows—where analysts typically juggle separate scripts, notebooks, and reports—with a single governed workbench. It ensures that critical risk metrics (like KS, AUC, and bad rates) are calculated by deterministic code rather than being guessed by an LLM, providing an audit trail for every decision.

How it works

Users describe their risk goals in natural language. The Agent then clarifies requirements, builds a validated workflow plan, and pauses at "responsibility gates" for human confirmation. Once approved, it executes deterministic tools to generate datasets, models, and reports. The system integrates several modules:

• Data Processing: Handles file registration, schema inference, and governed joins.
• Model Development: Manages recipe training, feature selection, and PMML export.
• Model Validation: Executes notebooks and compares scores to ensure consistency.
• Strategy Development: A seven-step process to design, evaluate, and deliver risk strategies (e.g., approval or pricing rules).
• Risk Analysis: Performs Vintage and roll-rate calculations.

Who it’s for

It is built for risk teams and credit analysts who need to develop governed, auditable risk models and strategies while maintaining human oversight and data locality.

Highlights

• Deterministic Execution: Uses platform code for calculations to ensure numerical accuracy and auditability.
• Local-First: Keeps files, task states, and outputs in a controlled local workspace.
• End-to-End Strategy Workflow: Covers everything from initial project understanding to the delivery of a final review package in multiple formats (Excel, Word, JSON, etc.).
• Human-in-the-Loop: Requires explicit human authority for high-impact actions and strategy adoption.