Work with me
What I build for you
I build agents, dashboards and pipelines for founders who want to know exactly what the machine did, what it cost and what changed. Everything waits for your yes.
I built five full sites for fictional brands so you can see the work before we talk. See the demos
Built for my own businesses first
I run these systems for my family’s companies and my own. A CRM that counts leads, deals, calls, mail, billing and Instagram posts. Storefronts with hardened checkout and subscriptions that renew themselves. A publishing pipeline where agents draft content as pending and the owner approves once in an admin queue. Nothing posts itself.
Seven machines, each agent with one job. A local media studio on hardware I own. A dashboard that watches every project: health checks, the queue of what only the owner can decide, reports.
- Agents draft in the CRM; a signed-in person presses Send or Publish.
- Custom storefronts for a coffee roastery and a hair-care shop, checkout hardened, subscriptions self-running.
- A media brand’s content pipeline: drafted by agents, approved by a human, posted only after that yes.
- 1,381 health checks across 28 projects in 20 days, run by the system, not by me.
One job, one closed list of permissions
Each agent has one job and a closed list of things it may touch. Research, drafting replies, reporting, the dull operational work that fills a week, fixing a project’s own failing checks. Anything outward, sending, posting, paying, deploying, shows you the exact text and waits for your yes. Single-use approval, expires in 30 minutes. Every time.
Checks are run by the system, not the model. Pass: the change goes in. Fail: it sits on a branch and lands in your queue. A suggestion that cites nothing never reaches you. Agents never deploy to production. Production is read-only to them.
- A daily cost cap you set, shown in every report.
- Models can run on hardware you own. Own the system, don’t rent it.
- 90 of 128 runs refused in one two-week window because unfinished work was already there. A gate that never says no is not a gate.
- Every fact comes with a source and a time.
Numbers, not adjectives
A weekly receipt: runs, failures, cost per run, hours saved. Every run and every check is written to a log you can read. Figures are read from the logs by a script, never typed by a person. If a log is missing or unreadable, the script stops and the page keeps the last counted numbers until the log is fixed; nobody fills the gap by hand. Each number says how it was counted.
The same reports track the business itself: leads, deals, content drafted versus approved versus published, search presence. You see whether it is growing week over week and keep going only if the numbers say so. One queue of everything only you can decide, so you are never asked in chat.
- Real numbers, 7–20 Oct 2026: 128 agent runs, 20 pushed clean, 9 held for review, 3 failed and flagged.
- Growth tracked in the same place: leads, deals, published content, week over week.
- A 30-day check-in after the first agent goes live, reading the numbers together.
Where your data lives and who touches it
Internal tools talk over a private encrypted network (a tunnel between machines, not the open internet). Keys live in one store on your side, referenced by name only, never in chat, logs, reports or my notes. Nothing goes out without your yes. Production is read-only to agents; a single write needs a human yes that day.
Backups get restored, not just taken. A real snapshot restored onto a separate machine, file count compared, files checked against the original. A 36-item security checklist is run before anything public ships. Your data is not used to train AI models and not sold.
- Proven restore: 74,616 files, 4.17 GB, restored and hash-checked 10 Oct 2026.
- I am building a decoy company with planted instructions and keys that unlock nothing, to learn what misbehaving agents try and set the guards against it.
- 36 commonly forgotten security items checked before launch: forged requests, exposed keys, unsigned webhooks, missing rate limits and more.
- Your keys stay in your store. I work with the names, never the values.
Start with the audit
The Agent Ops Audit at Integrated Autonomy. One week, fixed price, remote or in Los Angeles.
You keep the written map, one live agent, a one-page runbook and a 30-day check-in. After that: build more (fixed scope, fixed price, quoted before starting), then run (weekly receipt, monthly restore test).
- Days 1–2
Map a normal week and every job that repeats.
- Days 3–4
Pick the three jobs an agent can do today, what it does, what it costs to run, worst thing it could get wrong.
- Day 5
Ship one agent live behind your approval, with a log.