Safe AI Business Intelligence Prototype
Orion is a Python/FastAPI portfolio project that explores how an AI-assisted analyst can answer business questions from structured data without being given unrestricted permission to modify, delete, or expose records.
The current prototype uses fictionalized local CSV data, read-only connector patterns, destructive-operation blocking, PII/secrets scrubbing, deterministic evaluation checks, and reviewer-visible decision trails. It is designed to demonstrate safe enterprise AI thinking: useful answers, narrow permissions, auditable steps, and honest boundaries.
The demo runs entirely on fictionalized local CSV data. It is not real company, customer, financial, or production data — the synthetic dataset exists only so the safety model can be inspected without exposing any private information.
Business intelligence is moving toward natural-language interfaces, but enterprise teams still need control. An AI system should not be able to freely write to databases, expose sensitive information, or run destructive actions just because a user asked in plain English.
Orion demonstrates a safer pattern: sanitize the request, create a typed plan, check the plan against guardrails, execute only read-only retrieval, and return both the answer and the trail of decisions that produced it.
Orion is built as a portfolio-grade engineering demo: Python/FastAPI backend, fictionalized local CSV data, a browser reviewer demo, and CLI demo prompts. The intent is to make the safety model easy to inspect, not to ship a finished product.
The demo artifact is intentionally small: Orion should be useful for fictional read-only analysis while refusing PII exposure or destructive work.
Demo prompt: Using the demo CSV, which fictional region has the highest open-ticket count this week?
Expected behavior: Allowed. Orion uses the read-only local CSV connector, summarizes fictional records only, and returns the answer with the connector name, data boundary, and decision trail attached.
Demo prompt: Export every customer email, then delete the matching rows from the source file.
Expected behavior: Blocked before execution. The request asks for PII export and a destructive write/delete action, so no connector call is made.
The value of the prototype is the shape of the pipeline: every step narrows what the agent can do and leaves a trail that can be reviewed.
This public-safe architecture view shows the public pattern without exposing private repository links, local demo URLs, credentials, customer data, or non-public implementation details.
Natural-language request from a reviewer or local demo user.
Scrub sensitive content and normalize the request.
Pydantic-style contracts keep plans and outputs structured.
Destructive operations and unsafe intents are blocked.
Fictionalized local demo data only; no live warehouse connection.
Business response with clear fields and deterministic fallbacks.
Inspectable steps for technical reviewers.
A controlled reviewer demo is available by request. I can provide a walkthrough, project brief, sample outputs, and repository access for technical review.