An AI agent that writes credit-risk validation reports

What it does

MARVIS-Agent is a local-first web platform that uses an AI agent to guide credit-risk model validation. It executes validation notebooks, compares historical metrics through an agent memory foundation, and drafts structured Excel and Word reports from the results. The current stable release focuses on model validation, with broader modeling and strategy workflows planned for future versions. It treats governed financial work as something that should stay on local machines and leave a clear artifact trail.

The interesting bit

Most open-source AI agents chase coding assistants or chatbots; MARVIS chases the audit trail. Its core bet is that credit-risk work is too sensitive and regulated for cloud black boxes, so it keeps the agent, the notebooks, and the draft reports on your local machine. The Agent Memory Foundation tracks historical validation metrics so you can argue about model drift with actual evidence.

Key highlights

• Local-first execution via a FastAPI backend; no cloud required.
• Notebook-based validation runtime that generates reproducible artifacts.
• Agent mode drafts structured Excel and Word validation reports.
• Built-in Agent Memory Foundation tracks historical validation metrics for comparison.
• Private branding can be injected without modifying source code.

Caveats

• Only model validation is currently stable; broader modeling and strategy packs are on the V2+ roadmap.
• Local deployment is verified on macOS and Linux; Windows users face WSL2 path quirks.
• PMML scoring requires a separate Java runtime.

Verdict

If you validate credit-risk models and need local, auditable evidence, MARVIS is worth a look. Everyone else—including those seeking a general-purpose AI assistant—should keep scrolling.