AI Work

An LLM orchestration platform and two agents, all shipped by three engineers. I designed the Agent Builder and built its v1, owned the agent implementations and the frontend, and built an eval framework to measure agent quality.

Agent Builder

The Agent Builder is an LLM orchestration platform that lets you visually configure multi-agent workflows as state machines. It exposes each agent as a GraphQL + SSE API. We used it to deliver the AI products that follow.

Tap or hover over the numbered markers on the screenshots to learn more.

1The agent's state machine. Each node is either an LLM state (generate or chat) or a GraphQL state (for deterministic work without using tokens). Agent Memory drives the context and transitions between states. 2Configuration for the selected state. LLM states allow you to configure variables, the prompt template, LLM configuration, tools, agent memory, and state transitions. 3Chat preview allows you to test changes to the agent without deploying them, resulting in a fast feedback loop and confidence before shipping.

Purchase Order Agent

No need to manually process a purchase order with hundreds of line items. A swarm of parallel agents parses whatever the buyer uploads, verifies every line item against the live catalog, streams progress as it works, and hands ambiguous decisions back to a human before checkout. Built on the Agent Builder platform above.

1. Upload

1PO Agent handles whatever type of document the buyer provides, such as PDF, Excel, CSV, or photos, navigating vague product descriptions and messy handwriting.

2. Processing

1A swarm of parallel agents interprets the purchase order, verifies line items against the product catalog, and builds a checkout cart. 2Live progress streams to the browser over SSE as agents complete tool calls.

3. Review & Checkout

1Agent prompts user to make decisions for out-of-stock items, price mismatches, and other issues. 2Buyer responds in chat, and can request additional products or make other changes. 3Editable checkout cart. The buyer can override anything the agent did. 4Tax, shipping, and the buyer's saved address are pulled from their profile. Checkout is one click away.

Eval Framework

For the PO Agent, one-shot evals would only cover the happy path. The hard cases are purchase orders with real ambiguity and errors, where the agent has to resolve issues in conversation with the buyer. So every test case is a real PO plus an agent-simulated buyer and an expected cart, evaluated three ways.

1. Simulated buyer: An agent acts out the conversation with the PO Agent, following per-test markdown instructions (e.g. how to respond when asked about a product that isn't in the catalog). If the conversation goes wrong it fails fast and aborts the rest of the run.
2. Agent as judge: A second agent reads the finished transcript and scores how well the PO Agent communicated and resolved issues.
3. Hard assertions: Cart tools hit a mock, not the live API, and the final cart is asserted against the expected cart. Every PO has one correct cart, but many tool-call sequences are valid ways to get there.

Fitment Agent

An AI shopping assistant, embedded in the storefront, that finds products compatible with the customer's vehicle. Also built on Agent Builder.

1Embedded chat assistant helps customers find products compatible with their vehicle. 2Agent uses tools to search relevant vehicle data and renders interactive widgets to prompt user action. 3Fitment Agent is aware of the user's profile and selections made elsewhere on the site.