Skip to content
Home/Projects/Orion
Portfolio case study

Orion

Secure Agentic Business Intelligence Framework

A read-first AI business intelligence prototype that turns natural-language business questions into structured BI responses using FastAPI, LangGraph, Pydantic, fictional local CSV data, destructive-operation blocking, PII scrubbing, deterministic evals, and reviewer-visible decision trails.

Back to projects Private repo / available on request Live demo coming soon
Problem

Business answers without unrestricted data access

Orion explores how AI-assisted business intelligence can answer operational questions without giving an LLM unrestricted database access. The prototype treats every user request as something that must be sanitized, planned, checked, executed read-only, and explained.

Approach

Typed plans, guardrails, and inspectable trails

The system routes a request through sanitization, typed planning, guardrail checks, read-only connector execution, structured synthesis, and an inspectable decision trail. The emphasis is on narrow tool permissions, explicit intermediate state, and repeatable behavior that a reviewer can audit.

Build

Portfolio-grade engineering demo

It is built as a portfolio-grade engineering demo using fictional local data, deterministic fallback behavior, and repeatable build/eval checks. The public case study intentionally does not expose a running demo, customer data, live warehouse, ERP, CRM, or inventory connection.

Request path

A constrained BI agent pipeline

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.

01 Sanitize requestNormalize the question, scrub risky content, and prepare a bounded BI intent.
02 Plan with typesUse Pydantic contracts so plans and responses have explicit structure.
03 Check guardrailsBlock destructive operations and reject requests outside the demo's safe scope.
04 Execute read-onlyRun against fictional local CSV data through constrained connector logic.
05 Synthesize and show workReturn a structured BI response with reviewer-visible decision trails.
Sanitized architecture

A public diagram of the safe path, not private internals

This placeholder shows the public architecture pattern without exposing private repository links, local demo URLs, credentials, customer data, or non-public implementation details.

Input Business question

Natural-language request from a reviewer or local demo user.

Sanitize PII and risk screen

Scrub sensitive content and normalize the request.

Plan Typed BI intent

Pydantic-style contracts keep plans and outputs structured.

Guard Read-only gate

Destructive operations and unsafe intents are blocked.

Execute Local CSV connector

Fictional local demo data only; no live warehouse connection.

Synthesize Structured answer

Business response with clear fields and deterministic fallbacks.

Review Decision trail

Inspectable steps for technical reviewers.

What the case study demonstrates

  • How to separate natural-language BI requests from unrestricted database access.
  • How typed plans and structured responses make agent behavior easier to review.
  • How guardrail checks can block destructive or out-of-scope operations before execution.
  • How deterministic evals and fallback behavior support repeatable review.

Public boundaries

  • Prototype, not a production-ready product claim.
  • Uses fictional local CSV data, not real customer data.
  • No public live warehouse, ERP, CRM, or inventory-system integration is claimed.
  • The repository is private and the live demo is not exposed publicly yet.

Need a business system with clearer guardrails?

Bring the workflow and the risk constraints. The implementation should respect both.

Request a Local Tech Assessment