AI systems for owners and managers

AI that takes the manual work out of your orders, purchasing and data.

I'm Robert Sher. I build AI systems for distributors and product businesses, and I show the checks that prove they work before anyone relies on them.

Get the free one-page guideWatch episode 1
Cartoon of Robert Sher in a blazer pointing at an office easel
Your AI business solutions · episode 1

Marketplace orders arrive all day. Who types them in?

A 27-second look at the problem, what we built, and what we measured.

What we measured

Results, not promises.

~15 s · ~40 sto confirm and to enter a typical marketplace order. Of about nine recent orders, two waited for a person and the slowest took hours. A small sample, so a first look, not a promise.
16 of 16safety checks passed in one authorised real purchasing cycle, each seen to fail first.
97.6%of the reorder lines buyers went on to order kept the suggested quantity (1,096 of 1,123, over five weeks); 17% of suggestions were not ordered.
Common problems

Five problems, and what we built for each.

01

Orders typed in by hand

Built: an order assistant that reads each marketplace order and enters it into the ERP.

Measured: about 15 s to confirm and 40 s to enter, on about nine recent orders.

02

Reorders depend on one person

Built: a purchasing assistant that drafts the reorders; a person approves.

Measured: 16 of 16 safety checks, each seen to fail first.

03

The outside quote is a fortune

Built: a working prototype, put next to the quote.

Measured: not yet published; the quote-to-build comparison comes first.

04

Thousands of products, moving prices

Built: a store on a large product catalogue, fed by supplier price feeds.

Measured: live, selling through a major marketplace.

05

An old system nobody wants to touch

Built: a modern layer in front of the legacy ERP, one controlled way in.

Measured: tested end to end in a test environment.

Want all five on one page?

Plain English, one page, free.

Get the guide
Selected work

What it looks like when it ships.

Real screens from systems I built. Client names, prices and contact details are masked.

Product detail page with availability, dealer-benefit prompt and cart actions
Storefront home page with search, category navigation and sign-in
Category page listing bearing subcategories with product counts
Product detail with live availability
Industrial distributor · 2025 → now

Storefront on a large product catalogue

What it does: A headless B2B and B2C store with live availability, shipping rates and card checkout, fed by supplier feeds.

Result: Live, selling through a major marketplace.

PythonPostgreSQLStripeEDI X12
Arnold landing page
Live at arnoldtrainingapp.com
Own product · 2026 → now

Arnold, an AI strength coach

What it does: A coach that knows your training history, talks between sets, and tracks food and body state.

Result: On the App Store and Google Play, after working through several store rejections to a compliant build.

CapacitorRAGSwiftStripe
Pilot Reports dashboard listing filed capability gaps with severity, type and status, from Gap Hunter, Medic, Auditor, Reflector and Promoter
PilotGentic desktop app: radar of running agents, an embedded Claude Code session, a skills catalogue with per-tool toggles, and a copilot pane
Automation tab showing a recorded run: 58 actions, each with time, gesture, target app, coordinates and a screenshot thumbnail, with a Replay button
The PilotGentic Chrome extension overlaid on a web page: a stats panel with errors, console, network and performance, a status panel with a fix-this-page action, an embedded Claude Code chat, and a footer showing no issues
Gap report: 155 filed reports from the gap hunter, medic, auditor, reflector and promoter, each with a severity, a type and an open or fixed status.
Own product · 2026 → now

PilotGentic, a desktop for AI agents

What it does: Lets AI agents see and operate the Mac and browser, then audits every run and turns failures into reviewed fixes.

Result: Every run is logged and checked. 155 gaps filed and tracked to fixed or open.

MCPChrome extensionLLM evaluation
Survey welcome screen
Association name replaced
Trade association · 2026

A survey built for honest answers

What it does: AI-guided member survey where answers and booking contacts are kept in separate stores, so no one can be identified.

Result: Approved by the association task force for member distribution.

Privacy by designFastAPI
Home remodeling site with a free-estimate form
Public site with the estimate funnel
Home remodeler · 2025 → 2026

Chatbot, AI-assisted CMS and a zero-risk site move

What it does: A booking chatbot, a publishing system with SEO scoring, and a move to a faster static site with a revert snapshot taken first.

Result: 267 commits over a year; the live site was never at risk.

ExpressAstroLLM content
The platform behind it · one service per job
One front door screen, identifiers masked
One front doorSign-in to every internal tool
Marketplace EDI hub screen, identifiers masked
Marketplace EDI hubOrders, acknowledgements, invoices
RFQ automation screen, identifiers masked
RFQ automationQuotes drafted from email
Warehouse scanner screen, identifiers masked
Warehouse scannerReceiving, picking, counts

Full write-ups of each project are on the current homepage.

How we start

Start small, and prove it.

1. Pick one narrow jobFor example, quoting one product line.
2. A person approves every outputNo exceptions at the start.
3. Count what it got rightAnd what it missed. That count is your evidence.
Ready to move forward?

Get the one-page guide, then tell me what you would fix first.

Get the free guideEmail me