The numbers

Not the usual path

The usual route into AI work is a fine one. Tony took a different one, and the difference is the product. Here is the usual path beside what he actually has on the books.

Compare

The usual AI engineer

Tony

Where the skill came from

A degree, then a job building models or pipelines inside someone else's product. The customer is a ticket in a queue.

Years of making physical things for paying customers came first: CNC machining, laser cutting, resin, signage, commissioned pieces. Then came AI, learned by shipping it. He earned 37 certificates in eleven months and finished the last of them the same spring he built the first plugin.

What done means

The demo runs. The notebook executes top to bottom. Someone else carries the support inbox.

A reviewer approved it, a stranger paid for it, and the support page already answers their first question. That is 33 products live on the Elgato Marketplace, two more in Elgato review, a public support hub, and a public command explorer.

How he works with the model

An API call inside an application. The model is one component, and the conversation with it ends when the feature ships.

A working partnership, run like a production line. In one month he sat down with the model 363 times, every day of the month, and it carried out 36,654 actions on his instruction. Every session runs the same way across two Windows machines and a Mac, thanks to 10 written playbooks, seven purpose-built agents, and seven MCP servers he wrote himself.

When the software pushes back

File a ticket. Wait for the platform to add the feature. Ship the blind hotkey and call the gap a limitation.

Read the wall, find the door. His plugins reach 293 catalogued commands that their apps never gave a keyboard shortcut. Each one gets there by a route the app did not advertise, and every wall becomes a documented technique.

What the customer owns

The customer gets a subscription, a dashboard login, and a dependency on the vendor's roadmap.

The customer owns all of it. Plugins are one-time purchases on the Marketplace. Custom systems are handed over outright, source included. Nothing is locked in and nothing is rented back.

The evidence

A resume, a portfolio site, and a list of technologies, taken on trust.

The receipts carry dates. There is a public catalogue you can buy from, a command explorer you can search, 3,894 saved code changes since April across 47 repositories and one knowledge base, a support inbox with paying customers in it, and the numbers that follow.

The practice in numbers

What follows is five months of one person's output: the code, the products, the customers, and the hours spent with the model. All of it is measured, apart from one figure he reports himself, and he keeps revenue and spend private.

Five months of building

3,894

commits, or saved code changes, since 8 April 2026

About 26 a day, across 47 repositories and one knowledge base

868

commits in the busiest week

3 to 9 August. One every twelve minutes, all week long

104

work logs, 393,489 words recording what was decided and why

Each one kept up to date, never just added to

1,971

notes in his knowledge base

Searchable by meaning, not only by name

Commits per week, April to September 2026

Every repository plus his knowledge base, week by week. The last column is a partial week.

One plugin in April, a quiet May and June, and from July one shared engine grew into a product line. This is what organic growth looks like. Nothing was scheduled, and the work compounded because each product taught the next one.

Table view

Week of	Commits
Apr 6, 2026	78
Apr 13	65
Apr 20	23
Apr 27	3
May 4	0
May 11	3
May 18	0
May 25	0
Jun 1	0
Jun 8	0
Jun 15	0
Jun 22	0
Jun 29	10
Jul 6	241
Jul 13	131
Jul 20	333
Jul 27	554
Aug 3	868
Aug 10	551
Aug 17	594
Aug 24	343
Aug 31 (through Sep 2)	97
Total	3,894

One month with Claude

363

working sessions with Claude Code

3 August to 2 September. Nearly 12 a day

4,554

prompts sent to the model

About 150 a day. The AI carried out 36,654 actions in reply

995

tasks handed to helper agents

About 32 a day, and 171 multi-step jobs run start to finish

31 of 31

days in that month with at least one session

He sat down with the model every single day

What that month was, and what it built

97%

of what the model read was material it had already seen

Long, steady sessions rather than start-and-stop work

6.1billion

tokens, the model's units of text, through one desk in August

Four straight weeks over a billion, two Claude models in heavy use every week

3seats

paid Claude accounts, each run to its weekly usage limit

Nearly every week of August, by his own tally

Top 10%

of 1,080 Claude Code developers on a public usage leaderboard

Ranked by monthly usage, counting one of his three accounts

products live on the Marketplace, every one through Elgato review

30 paid, 3 free. The newest published 2 September

447

plugin installs in the 30 days to 28 August

Without a single ad, post, or promotion behind them

customers who wrote in unprompted, over 30 days

The UK, South Korea, and the US. All found the catalogue on their own

10,000+

pageviews on the company website by 5 August

Three weeks after the site went live

All of that work is on the shelf where a customer can see it. It became products that sell without a salesperson, a support record any client can read, and 35 worked examples of reaching software that resists automation.

Built to work this way

MCP servers written from scratch, connecting the model to his tools

Windows automation, a two-machine bridge, a live terminal, LinkedIn, and more

written playbooks for the model, plus seven purpose-built agents

2,672 lines of written instruction, all his own

MCP servers in his working setup

Seven he wrote himself, four from other makers

machines driven from one chair

A desktop, a laptop, and a Mac mini

37 certificates, in the order they were earned

Coursera credentials by issuer and month, June 2025 to April 2026. Larger dots mean more credentials that month.

Read it left to right and the plan shows. Model theory came first, 16 IBM courses deep. Automation engineering with Google came next, so the theory could run unattended. Formal manufacturing method from Arizona State and Buffalo put a framework under the years at the machine. He finished at Anthropic, on the tools he builds agents with.

Table view

Issuer	Jun 25	Jul	Sep	Oct	Nov	Dec	Jan 26	Feb	Mar	Apr	Total
IBM	3	1	1		2	10					17
Google		1		1	1		2	3			8
Arizona State								8			8
University at Buffalo								1			1
Anthropic									1	2	3

How 2,938 catalogued commands reach their applications

Drawn from the public command explorer, which covers 27 plugins.

Keystroke the app already understands · 2,645Reached by another route · 293

Ten percent of the catalogue is the part that took the thinking. These are commands the apps never gave a keyboard shortcut. The plugins reach them anyway, by name over a command line, through the app's own API, over a network message such as UDP or OSC, by driving the interface directly, or by installing the shortcut the app forgot to ship. A deck that only sends keystrokes stops at the other ninety percent.

The full line's total of 3,914 commands includes plugins not yet in the explorer. Search the explorer.

Table view

Route	Commands	Share
Keystroke the app already understands	2,645	90%
Reached by another route	293	10%
Total	2,938	100%