Skip to content
Subscribe

Mike Hodgen · San Diego

I build AI that runs companies.

Starting with my own: a vertically integrated manufacturer and apparel brand, where agents do the work and I make the calls that matter.

Updated Nov 1700+ AI agents and automations600+ evals that grade AI output100+ AI models from 14 companies

I started that company with $200 of fabric and built it into printing, sewing, selling and the books under one roof. Now it runs on AI agents I built. I'm at my best on ideas and hard problems, and I've done most of the jobs in a company with my own hands.

Mike Hodgen

The range

Seven disciplines. All hands-on.

Most people who talk about AI have touched one corner of a business. I've worked the factory floor, the checkout, the ad budget, the payroll and the agents myself, so the problems connect for me.

  1. ManufacturingUV printing, cutting, sewing and shipping, all under one roof I set up.6 steps, one buildingThe factory
  2. Supply and inventoryForecasting what to make, and freeing cash that sat on shelves.$300K → $100KOpen the case
  3. Ecommerce and checkoutMy own storefront, catalog and checkout, and the billing logic that decides who actually paid.My own checkout and billingRead the note
  4. GrowthPaid ads, email and texts, search, and showing up in AI answers.2.39x return, brand defense keptOpen the case
  5. Finance and payrollThe books, the monthly close, cash forecasts and a payroll engine that follows California wage law.29 of 29 law examplesOpen the case
  6. Legal and complianceRecords that hold up later, and products rebuilt when the law changed under them.Void, never deleteRead the note
  7. AI engineeringAgents, the checks that catch their mistakes, and the judgment of when to keep a person in the loop.500+ AI agentsOpen the case
  • Illustration of a UV printer head printing a stained-glass pattern onto white stretch fabricIllustration
    Print
  • Illustration of garment pattern pieces cut from printed fabric on a digital cutting tableIllustration
    Cut
  • Illustration of an industrial sewing needle joining butterfly-print fabric to black meshIllustration
    Sew

Illustrations

The casebook

Problems I've solved. Open one.

Each one starts with what went wrong, then the part most people miss, what I built, and what happened. Real numbers only.

Case 01Growth · EmailMy emails were landing in spam, and every authentication check said they were fine.
  1. The insight

    Authentication was never the problem. The open-tracking pixel was.

  2. What I built

    I tested the cause three independent ways before touching anything, then turned the tracking off.

  3. The result

    Deliverability testers scored the email 8.9 and 9.5 out of 10, and the domain was clean on 22 of 23 blocklists.

Read the whole case
Case 02AI engineeringAutonomous systems kept reporting success. A data pipeline I inherited had been dead for ten days behind a quiet error.
  1. The insight

    AI rarely fails by doing the wrong thing. It fails silently and reports what it meant to do.

  2. What I built

    Every action is read back from the real system before it counts, and monitors look for fresh timestamps over empty data.

  3. The result

    It is now a rule in every AI system I ship, and I have shipped more than 15 into production.

Read the whole case
Case 03Growth · Paid adsI gave AI the ad budget. On paper, the weakest campaign was the one bidding on my own brand name.
  1. The insight

    Some spend looks bad on its own because it protects everything else.

  2. What I built

    Guardrails the AI cannot override, starting with one rule: brand search is untouchable.

  3. The result

    The account returned 2.39x overall while brand defense sat at 1.07x, and stayed funded.

Read the whole case
Case 04Finance · PayrollPayroll tools are built for the federal weekly rule. California pays overtime by the day, with meal premiums on top.
  1. The insight

    A $3 mistake on a meal premium, repeated across people and pay periods, becomes a six-figure liability.

  2. What I built

    A payroll engine that encodes the wage law itself, tested against the published court and agency examples.

  3. The result

    It passes all 29 worked examples, including daily overtime, meal premiums and bonus rules.

Read the whole case
Case 05Supply · InventoryCash was tied up in stock on the shelves while the business needed it to grow.
  1. The insight

    The fix was the forecast and the operating decisions behind it, not more software.

  2. What I built

    Better forecasting with a planning tool, and different calls on what to make and when.

  3. The result

    Inventory held went from about $300K to about $100K at its lowest.

Read the whole case

Outcomes

What it did for the business.

Scope: one company I operate, a vertically integrated manufacturer and apparel brand. These are its published results, each with its period and source.

All published results
Revenue per employee, 2024. First-party.
+38%
Less manual operations time, 2024. First-party.
42%
Cash in inventory at its lowest, about $300K before. First-party.
~$100K

How I think

Five rules, learned the expensive way.

Agents do the work I make the calls
  1. 01

    Agents propose. I approve.

    Autonomy is earned one task type at a time, never handed out. Why

  2. 02

    Make the failure mode "nothing happens".

    Every risky system ships switched off, with a kill switch I can reach. Why

  3. 03

    Read it back before you believe it.

    A system that reports its intentions will lie to you with a straight face. Why

  4. 04

    Ground it in the real data.

    Answers are locked to the real catalog, prices and books, so nothing gets made up.

  5. 05

    Say what doesn't work yet.

    It builds more trust than any claim, and it tells me what to build next.

The ledger

Updated November 1, 2026

Code changes in 2026, through November 1, 2026

23,722

October: 7,355, about 5 times March, the first full month.

  1. 1,419Mar
  2. 1,386Apr
  3. 1,427May
  4. 1,383Jun
  5. 2,655Jul
  6. 3,292Aug
  7. 4,238Sep
  8. 7,355Oct
Full months only, from March. Red is the latest.
Updates shipped live
3,636
AI agents and automations
700+
Software projects
109
Of my code written with AI
About half
AI models built with
100+
Departments run by agents
7
Skills I use and build
150+
Evals: tests that grade AI output
600+
Automated checks and release gates
1,000+
How I count

The full method, including what I leave out and the code that deduplicates: How I count the ledger

  • Code changes. Unique code changes in 2026 across every local codebase and branch, each counted once by its fingerprint. The total runs through the update date; the chart shows full months only.
  • Updates shipped live. Changes merged into live software since January 1, 2026, from GitHub.
  • AI agents and automations. Distinct AI agents with their own role across every project, plus jobs that run on their own schedule. Counted once each; copies and plain code are excluded. That is 535 agents, shown as 500+, and 744 with the scheduled jobs, shown as 700+.
  • Updated. I refresh these numbers from my own history and GitHub; the date above is the last refresh.
  • AI models. Distinct AI models I have built with across current and retired projects, from 14 companies. Products are named by model family (Seedance is ByteDance, Kling is Kuaishou, Wan is Alibaba, Hailuo is MiniMax, Flux is Black Forest Labs).
  • Skills, evals and checks. Reusable skills I use and build for my agents (165 unique, many installed rather than written by me). Evals: tests that grade AI output, about 650 across 32 projects. Automated checks and release gates (replays, scorecards, page-release gates), about 1,700 counted and rounded down.

Meet the crew

120 named agents. A few of them.

Every one has a job, a name and limits I set. Some are private, so these are from the 74 I can show you.

  • Rich

    Sorts my email every 5 minutes, up to 50 messages a run, then writes a morning brief.

    Chief of staffRunning

  • Suzy

    Keeps the mail a real person sent and archives the rest before Rich sees it.

    Chief of staffRunning

  • Aura

    Talks to real customers, and never issues store credit over $25 on its own.

    Store and supportRunning

  • Maestro

    Plans the month and proposes one campaign a day. By design, it sends nothing.

    Growth OSRunning

  • Megatron

    Runs twice a day beside my social media manager, who picked the name.

    The factoryRunning

  • Ava Sterling

    Directs a 12-agent crew that turns an approved caption into a 10-second narrated reel.

    Hollywood StudioRunning

  • Critic

    Fails closed: if it can't judge a campaign brief, a person does.

    Growth OSRunning

  • Muse

    Push to talk. Scans trends, finds gaps in the catalog and queues new designs for my approval.

    My own design agent, not Meta's ad product of the same name.

    The factoryRunning

Meet all 74, grouped by the system they work in

What I work with

100+ AI models from 14 companies, built with by hand:

  • OpenAI
  • Anthropic
  • Google
  • xAI
  • Perplexity
  • Voyage
  • Cohere
  • Seedance
  • Kling
  • Wan
  • Hailuo
  • Flux
  • Midjourney
  • ElevenLabs

And the tools I ship with every day:

  • Next.js
  • Supabase
  • Vercel
  • Stripe
  • GitHub
  • Resend
  • Claude Code
  • Codex

Follow the build

I write down what I learn while it's still fresh.

What I built, what broke, what it cost, and what I'd do again. Sent when there is something worth your time.