Self-built product

Glass Box

An AI agent system with three agents for support tickets, supplier invoices and business questions, where anything risky waits for a person to approve it.

A complete product I designed, built and tested myself. Client work is in Case studies.

Glass Box live console pausing a support agent's refund over the $150 threshold until a person approves or rejects it, with the agent's workflow path below

Agents that act need someone in charge

An agent that can refund a customer or pay a supplier is only useful if a business can trust it. That means knowing what it did and why, stopping it before it does something it shouldn’t, and knowing what each run costs.

Three agents, with a person approving anything risky

Glass Box runs three agents: one handles support tickets, one matches supplier invoices against purchase orders, and one answers business questions with live SQL. Anything risky waits for a person. A refund over the limit, an invoice that doesn’t match its purchase order, or a report for leadership pauses the run until someone approves or rejects it, and the agent resumes where it stopped.

A control room traces every step, with the agent’s reasoning and what the step cost.

The rules live in code, not the prompt

Glass Box is built in Python with LangGraph behind a FastAPI service. Its rules live in code rather than in the prompt: card numbers are masked before any model sees them, prompt-injection emails are quarantined with every write action switched off, and step and cost caps stop runaway runs.

Tech stack

  • Python
  • LangGraph
  • FastAPI

A regression caught before it shipped

A 180-case evaluation suite caught a prompt change that broke policy citations and skipped refund approvals. Smarter model routing cut the cost per resolved ticket by about a third. 96 automated tests cover the system.

Screens from Glass Box

Glass Box run trace: the workflow path from intake guard to reply, timed steps including PII redaction and a prompt-injection scan, and the policies behind the agent's refund decision

A run traced step by step, with the agent’s reasoning.

Glass Box evaluations comparing three versions of the support agent: version 13 dropped to 85 of 120 cases with refund approvals skipped, and version 14 restored 119 of 120

The evaluation suite comparing versions of the support agent.

Glass Box cost and models page: spend per day by model, cost per resolved ticket by agent version, and a tool to re-price a step on another model

Cost by model and per resolved ticket.

Want something like this for your business?

Tell me what you need it to do. I’ll reply within 24 hours on weekdays with questions or a suggested first step.

Start with the $150 scoping session, credited toward the build.