Kellen Poole ← All projects

Agentic AI · resilience engineering

ShopFlow-Orchestrator

Turns raw technician inspection shorthand into a priced repair order, an empathetic customer text, and an advisor cheat sheet — with a self-correcting resilience layer that still ships an estimate when the vendor API and the AI are both offline.

Live output · 30/30 tests pass

Started June 2026 · self-directed

See live output → Source code ↗

The problem

There's a bottleneck between a technician finishing an inspection and a customer approving repairs: terse shorthand has to become a priced, clear, empathetic set of outputs — and it has to keep working when a parts-vendor API or the AI is down.

How it works

  1. Parse technician shorthand (e.g. "front lower CA bushings torn, play in ball joint, rack & pinion weeping") into structured repair items.
  2. Price each item — parts, labor, and margin — via an underwriting layer.
  3. Produce three outputs: a customer SMS/web notification in plain English, an internal priced repair order, and an advisor cheat sheet with talking points and objection handling.

The story · STAR

S

Situation

The gap between a technician finishing an inspection and a customer approving repairs is messy — shorthand has to become priced, empathetic outputs, and it has to survive a vendor API or AI outage.

T

Task

Build a pipeline that converts technician notes into customer-ready summaries, advisor talking points, and a priced order — with autonomous self-correction.

A

Action

Built a zero-dependency Node pipeline (five stages) with a circuit breaker, exponential backoff, an emergency-pricing matrix, and layered fallbacks. The LLM path is optional behind an API key, with a deterministic domain simulation otherwise. Added a human-review queue for degraded runs and 30 tests with fault injection.

R

Result

A fully runnable, tested system (30/30 tests). Even with the parts vendor down and the AI offline, it falls back to emergency pricing + template copy and still ships a complete estimate — auto-flagged for human review.

Watch it run · 30-second demo

A real run end-to-end: the technician's shorthand becoming a priced estimate, the customer text, the internal repair order and advisor cheat sheet, and the degraded-mode resilience fallback that still ships an estimate.

Open the live output →

Built with

Node.js (zero-dep)circuit breaker + fallbacksoptional Anthropic SDKnode:test · 30 tests

Status & honest limitations

Fully runnable & tested (30/30). Zero external production dependencies.

  • The real LLM path is optional behind an API key; without it, a deterministic domain simulation runs.
  • A personal proof-of-concept — not deployed in a real shop.

Self-directed project, built with AI-assisted development. All figures shown are illustrative.