Bring me a task your team repeats every week. I work out what to automate, what still needs a person, and what a useful first build looks like. Then I build it and keep it running.
I use AI where it helps and plain automation where that is enough. Everything I build has checks around it: source checks, a separate review, and a person's approval before anything sensitive happens.
What clients say
Websites & AI consulting
Websites
“Julian at Metra built the Just Results Personal Training website quickly and thoughtfully. He captured our brand and made it easy for people in our community to find us. …”
Trainer Joe, Just Results Personal TrainingI rebuilt the Just Results website, added search optimization and created custom business cards.
“Julian helped out so much with his website, I have gotten so much more business from his website and the customers love how easy it is to navigate through the website and have no issues with trying to contact me. Overall 5 stars with the website that julian built for me.”
Francisco, Precision Mobile MechanicI built and maintain the Precision Mobile Mechanic website, with clear ways for customers to call or text.
“Julian is the man he got my website up and going everything he promised me he did any questions or concerns I had he answered anything I want to change on my website or anything i want removed from it he's been on top of it usually getting it done within a couple hours at the most if you have any doubts of having him get your websites done don't hesitate just have him do it he will not disappoint”
Martin, 303 Rapid RecoveryI built the 303 Rapid Recovery website and redesigned its homepage.
AI consulting
“Had an amazing call with Julian here helping me set things up with Claude, linking LLMs together and building bots for my business. …”
Ishaan KumarI helped Ishaan set up Claude, connect language models and build bots for his business.
“pay for some consultancy to help get up to speed quickly because one hour of his time has effectively bought me back hours a day”
Ishaan KumarI helped Ishaan set up Claude workflows, reports and lead lists.
Latest work
Picked and updated automatically
The strongest recent things I have built. This list updates itself every week: each website in it is opened and screenshotted again, and the picks are made again from finished work.
Built the shop's website, then redesigned its homepage through his site factory; the owner picked design B.
The factory design B went live on October 23, 2026.
It serves an llms.txt file for AI answer engines.
Screenshots taken November 16, 2026
Kind
Tool
Status
Finished
Premium site factory with award judge
Built a repeatable workflow for premium websites: it reads a client's current site, studies award-winning sites, builds three homepage designs, tests them, and the client picks one.
303 Rapid Recovery's live homepage came out of it.
Built a new auto-repair company's website with its own look, service pages, Spanish wording and visit/call tracking.
camelautomotive.net went live on October 21, 2026.
A windshields page and Spanish wording were added on October 24, 2026.
Screenshots taken November 16, 2026
Kind
Tool
Status
Finished
Approve-then-post tool for X
Built a tool where AI drafts posts, he taps Approve on his phone, and a scheduled function posts through the official X API.
It only posts through the official API and never auto-replies, likes or follows.
It posts at most 3 times a day.
Kind
Tool
Status
Finished
Flyer Map app
Built a phone map app where he taps the side of a street to mark flyer drops done, synced to a server.
Version 1 used per-house GPS taps and was scrapped because GPS put drops one house over.
Version 2 went live on October 8, 2026.
Website case studies
Results checked October 17, 2026
Two local businesses, the websites I built for them, and what the search data shows. Each case study includes the dates, the full comparison and how the numbers were measured.
Client
Precision Mobile Mechanic
Work
Website and local SEO
Location
Thornton & the Denver metro
Measured
September 18–October 15, 2026
Sources
Google Search Console & Google Analytics
Helping drivers find a mechanic who comes to them
A website organized around repairs and the cities Precision serves, with clear call and text options. Unpaid search accounted for about 58% of its recorded website visits in the latest 28-day period.
135Clicks from Google Search
219Website visits from unpaid search
Some search positions improved, while total traffic fell compared with the previous period. The case study shows both.
“Julian helped out so much with his website, I have gotten so much more business from his website and the customers love how easy it is to navigate through the website and have no issues with trying to contact me. Overall 5 stars with the website that julian built for me.”
Francisco, Precision Mobile MechanicI built and maintain the Precision Mobile Mechanic website, with clear ways for customers to call or text.
Client
Able Towing of Cortez
Work
Website and search visibility
Location
Cortez, Colorado
Compared
August 21–September 17 vs. September 18–October 15, 2026
Source
Google Analytics
Towing, repairs and truck parts. One clear place to find them.
Service sections, real fleet and shop photography, and a direct route to calling. The latest 28 days show early growth in search traffic, from a small starting point.
19 → 27Visits from unpaid search · up 42%
39 → 56Total website visits · up 44%
Eight more visits from unpaid search. These are recorded website visits; confirmed calls, jobs and revenue were not available.
I designed, built and run all six. The case-study pictures use sample names and data. The Matrix image is a recreated picture of the setup, with private names and text replaced. The numbers in the text come from the systems themselves.
Kind
AI workflow software
Status
My daily working setup
Structure
One CEO, one COO and six workers
Heavy work
A second computer handles builds and long tests
One conversation. A coordinated team of AI workers.
Matrix is the workspace I built to run my work with AI. I talk to a CEO agent in one window. A COO agent organizes the work and assigns it to six worker agents, each with one named job and a visible status card.
Workers bring anything that needs me back up as one short question. Sending messages, spending money and publishing social posts require my exact approval. Larger builds and long tests run on a Mac Mini, so they do not take over the computer I am using.
Before a window runs out of working memory, its progress is saved and handed to a fresh session. The next worker picks up the same job with its decisions and limits intact.
Picture of the setup, recreated from the working layout and task cards. Private names and text replaced with neutral labels. Not a screenshot. Select the picture to view it full size.
Kind
AI automation
Status
The form is live. The automatic first draft is built and tested end to end on sample requests, and is being switched on.
A client describes the website they want. The first draft builds itself.
My site builder makes a demo website from a business's public Google listing, with no form needed. When a client does fill in the form, with their photos, their logo and a written description, what they put on it wins.
One AI step turns their words into a short, checked brief. The builder applies it as a first draft and texts me what was done and what still needs a person.
The AI only fills in a brief. It can't touch the site. Colors must be real color codes, text has length limits, and anything that reads like an instruction or a link is thrown out.
Case study, 2 pages. Select a page to read it full size. Download the PDF
Kind
Data funnel with AI checks
Status
Ran in September 2026 for a white-label postcard campaign.
Numbers
13,812 businesses checked in seven Colorado regions
1,149 AI website checks, each re-tested by code
460 qualified leads, 423 of them mailed
About $0.30 in data costs per qualified lead
Built with
Python, Google Places, Colorado's public business filings, address verification, Claude
13,812 map listings in. 460 businesses worth a postcard out.
A nine-step data funnel that finds local businesses with no working website, and the owner to address a card to. It checks a whole region the same way every time.
AI makes the one judgment call a script can't: does this business have a website its listing doesn't show? A separate code check then visits every site the AI reports.
Finding the owner. The shop's address is checked with the mail service before any card is paid for. Then the business is looked up in Colorado's public business filings to find the registered agent, who is usually the owner, and the address the state has on file for that person. The card is addressed to a named person at that address, which is often not the shop, so it lands with the owner instead of the front counter.
When the AI isn't sure, the answer stays "not sure". Every paid lookup is logged before it is sent, so a re-run never pays twice. Every business that drops out carries the step and the reason.
Case study, 2 pages. Select a page to read it full size. Download the PDF
Kind
Direct mail automation
Status
The 423-card run mailed on October 3, 2026. Results are not in yet.
Numbers
576 personalized cards mailed, July to October 2026
423 in one white-label run for two agencies
Under $1 per card, printed and mailed
5 minutes between scan checks, all day
Built with
Python, headless Chrome, the Lob print-and-mail API, Netlify, text alerts
A postcard that shows each business its own new website.
Every card carries a picture of a website already built for that exact business, and a QR code that opens it. The cards are made, checked and mailed by code. When someone scans one, I get a text within minutes, while they are still looking.
It runs white-label. Two agencies' cards went out in each agency's own name, colors and wording from the same system, and each agency's links can only open that agency's demos.
Case study, 2 pages. Select a page to read it full size. Download the PDF
Kind
Monitoring and auto-healing
Status
Runs today on nine live client websites, their contact forms and my background jobs.
Numbers
Every 2 minutes: each live site is loaded and checked
Every 15 minutes: each contact form is checked
3 tries, then it stops and texts a person
6 repairs a day at most for any one system
Built with
Python, scheduled jobs on a Linux server, the hosting and email providers' APIs, text alerts
A watchdog that fixes what's safe to fix and only calls a person for the rest.
Most monitoring just sends alerts. This one checks a system the way a user would, makes sure a failure is real, rules out its own blind spots, returns the system to its last known-good state, and proves the repair worked. It stays quiet unless a person is needed.
It never acts on a request that got no answer. It can't loop. It only ever goes back to a state that was live and passing, and a change a person made is never overwritten.
Case study, 2 pages. Select a page to read it full size. Download the PDF
Kind
Email automation
Status
The client record runs on its own. Sorting has run by hand while its schedule is switched on. The reply assistant has run since August 2026 on my own outreach, which is now paused, so no real reply has needed a draft yet.
Numbers
3 times a day: a client's email record saves itself
131 automated tests on the reply assistant
Built with
Python, the Gmail API, Claude, a calendar API, Netlify Functions
My inbox sorts itself, saves client files, and drafts replies for my OK.
Rules file my email. Bills and notifications are filed away, client mail stays in front of me, and anything no rule matches is left alone. A client's whole email history saves itself as one transcript, with every attachment.
For my outreach, AI sorts each reply and drafts the common ones in wording I approved in advance, with two real open times from my calendar.
Nothing sends until I approve it from my phone. The AI sorts replies. It doesn't write freely, and when it isn't sure, it doesn't draft.
Case study, 2 pages. Select a page to read it full size. Download the PDF
Also built
AI agent workflows with review and permission gates
My daily setup runs several AI coding agents in isolated workspaces. A separate review checks each build, and the agents can't send, spend or delete without my approval.
Technical write-ups, one page each
The dense versions, for a developer or a hiring manager who wants the detail.
Automated site generation and deployment pipeline.One page. Select it to read it full size. Download the PDF
Outbound email pipeline and deliverability engineering.One page. Select it to read it full size. Download the PDF
How I work
Checks around the work
Source checks drop any claim that can't be traced to a real record. A separate review looks at each build. Sending, spending and deleting wait for a person's approval.
AI where it helps
I use AI for the judgment calls a script can't make, and plain rules everywhere else. Sometimes the right answer is a tool you already own or a simple rule, and I'll say so.
Honest status
I say what is built, what is switched on, and what hasn't produced results yet. You can see that in every case study above.
Tested by breaking it
A safety rule only counts if a test fails when the rule is removed. I break each one on purpose to make sure the test is real.
Ways to work with me
A project
A website, an automation or an internal tool. A fixed setup fee plus monthly support, agreed before I start.
White-label for agencies
I build under your brand. You keep the client relationship, and we agree up front how client contact works. You work directly with me, the person doing the work.
AI consulting
$100 per hour. The first 20-minute call is free. Bring one repeated task. We work out what to automate, what needs a person, and what a useful first build would be.
Contract or part-time
I'm open to contract work and part-time roles, remote or around Denver and Boulder.
AI agents
Workflow automation
Software development
Generative AI
Data pipelines
Business process automation
Web development
Web design
SEO
Google Ads & Meta Ads
CRM
Marketing automation
Prompt engineering
REST APIs
Software testing
Get in touch
The fastest way to reach me is a LinkedIn message. Tell me the task or the project in a sentence or two.