SS
All work

Mar 2026 – Present

BayOps AI

Voice-Driven Parts Ordering for Auto Repair Shops

Service advisors lose hours on multi-vendor parts ordering, and LLM-driven checkout flows quietly invent prices and order parts that don't fit the vehicle.

I don't trust LLM output by default — every money-touching step passes a deterministic validation gate before it hits a real cart.

View source

Pipeline

End-to-end flow, stage by stage.

  1. 01

    Voice intake

  2. 02

    Vehicle lookup

  3. 03

    Fitment check

  4. 04

    Price sanity gate

  5. 05

    Browser checkout

  6. 06

    Excel billing sync

How it works

01

Voice in from the bay or over the phone

Speech arrives through the browser mic, an uploaded file, or an inbound Twilio call to the shop's number. Both paths run the same pipeline, and ElevenLabs speaks replies back so the mechanic never has to look at a screen.

02

Fitment verification before pricing

A verification engine combines NHTSA vehicle data with a Groq-hosted model to confirm the part actually fits the stated year, make and model, and the assistant asks for whatever detail is missing instead of assuming.

03

Parallel multi-vendor price search

Every part is searched simultaneously across AutoZone, NAPA and Advance Auto. A single batched Groq call extracts prices for one part across all three vendors, results cache for 10 minutes to avoid repeat search hits, and a sanity check rejects hallucinated $0.01 or $9999 prices.

04

Cart automation and live estimate

A Playwright browser agent adds the cheapest verified part to the vendor cart and re-checks it afterwards, while the estimate updates live with markup and tax, supports chat price overrides such as "set brake pads to $45", and exports to Excel per order.

Engineering decisions

  • Built a voice-driven ordering assistant that takes a shop's request end to end, across multiple parts vendors.
  • Ran parallel searches across AutoZone, NAPA, and Advance Auto, using a 10-minute search-session cache and batched Groq extraction to reduce repeated lookup work.
  • Eliminated hallucinated-price checkout errors with a price-sanity validation layer between the model and the cart.
  • Verified part fitment in real time by combining NHTSA vPIC API vehicle metadata with LLM reasoning.
  • Automated vendor checkout with Playwright and Claude Computer Use, driving real browser carts.
  • Synced live billing and markup estimates straight into the shop's Excel workbook via xlwings.

By the numbers

Pipeline efficiency

Batched extraction and parallel search against a naive per-vendor loop.

Vendors priced per part

higher is better
Before
1
After
3

LLM extraction calls per part

lower is better
Before
3
After
1

Repeat vendor hits within 10 min

lower is better
Before
Every query
After
Cached

Ordering channels supported

higher is better
Before
1
After
3

What it ships with

  • Twilio phone-call ordering with signature validation
  • Text chat advisor alongside voice
  • 10-minute search cache to cut vendor lookups
  • Manual price override from chat
  • Post-checkout cart verification
  • Live Excel export per order

Outcomes

0
Hallucinated-price checkout errors
Multi-vendor
Ordering from one voice flow
Real-time
Fitment verification via vPIC
Live
Excel billing sync

Stack

PythonFastAPIReact 19GroqElevenLabsLiveKitTwilioPlaywrightClaude Computer UseNHTSA vPIC APIxlwingsLLM Agents

Next case study

RACE

RAG-Optimized Clinical Reasoning Engine

TruckerPath

Fleet Operations & Intelligence Platform