← Back to Projects 01_PROJECT // ANVIL

Anvil

Active

AI-Empowered Repair-Tech Platform · Mindlink Research Group LLC

02_OVERVIEW

What is Anvil?

Anvil turns the diagnostic catalog of a bench repair technician — a Best Buy Advanced Repair Agent, or any in-shop tech — into a small AI agent that runs on the technician's phone. The agent retrieves the right diagnostic skill, calls structured tools, and walks the tech through remediation, all while tracking cost per case in real time.

It is built on a deliberately model-agnostic architecture: the value lives in a curated library of Markdown skill files and a published PowerShell tool contract, so the underlying model can be swapped or downgraded without losing capability.

03_DESIGN_CONVICTIONS

Three Convictions

01

The skill library is the product

Markdown skill files are the asset; the model is replaceable infrastructure.

02

Small model + right context wins

A small model with the right context beats a frontier model with none. Tiered routing (Haiku → Sonnet → Opus) keeps cost-per-case in the cents.

03

PowerShell is the action layer

Structured JSON in/out under a published contract. The model never reads raw logs.

04_UNIT_ECONOMICS

Validated Numbers

$0.04
AI cost per real bench case
$0.05
AI cost per consumer support session
87.9%
Gross margin per Anvil Assist sub ($9.99/mo)
$110/mo
Pilot AI cost (5 stores, ~2,600 sessions/mo)

Figures from a token cost model with an independent math check; tiered routing keeps per-case AI spend in the cents.

05_STATUS

Current State

A runnable Python CLI alpha ships today: 10 skills across diagnostics, remediation, and client-comms; five PowerShell tools wired end-to-end; mock fixtures for three demo scenarios (BitLocker recovery, slow-boot SMART regression, BSOD post-update); and 31 passing tests that require no API key. The next milestone is a field-test proof of concept measuring real cost-per-case across live bench tickets.

06_STACK

Tech Stack

Python Anthropic Claude Tiered Routing PowerShell BM25 Retrieval JSONL Audit Rich REPL