Agents that finish the job.
Multi-step work with tools: reading files, calling your systems, making things and checking the results. The model fills in a structure and code checks it before anything runs.
AI engineering · Phoenix, Arizona
I build AI that does real work.
Agents that finish multi-step jobs. AI features inside the software you already run. Models on your own machines when your data can't leave. I'm Vincent Serpico, and I've been writing software for more than 30 years.
Agents and tool useAI features in your softwareLocal and private modelsTested and measured
What I build
Three kinds of
AI work.
A model on its own makes mistakes. The engineering around it decides whether those mistakes reach you. That part is my work.
AI inside your software.
Assistants, search, drafting, extraction and generation, built into the product or internal tools you already have. Usage is metered and limited where it costs money.
Private AI on your machines.
Models that run on hardware you control, for data that can't go to a public AI service. Set up, connected to your systems and tested on the real task.
Inside Prompt Maker
How Prompt Maker's agent turns one sentence into a short film.
Prompt Maker's assistant runs on a small model on the user's own computer. Small models drift, so the app owns the structure and the model fills it in. Here is the path one request takes. Select a step.
STEP 01 · PERSON
Ask
The result
One fisherman, four scenes, 22 seconds.
Mending his net at dawn, laughing over cards in a harbor bar, steering through a storm, walking home at night. The same face and sweater in every shot. Made on one home computer with free models.
Guardrails
Deleting, moving or overwriting a file waits for the user's click, however the command is written. A third video render in one chat turn asks first. The agent stops itself after asking the app the same thing three times, and it can't approve a cloud model on the user's behalf.
Context
Before each turn it checks how much room the model has. Older steps shrink to one line, then fold into a short progress note. The app keeps its own record of what every tool reported, so finished work isn't redone.
Evals
An eval types real requests into the assistant and scores each local model: how often its plan is right the first time, how often it's right after the app sends it back, and how long it takes.
brain-eval.mjs ↗Tests
The whole path runs in headless Chrome against a scripted model and a fake ComfyUI, with real mouse clicks, failing on any console error. A change to a check or the compiler is caught without a GPU.
ui.test.mjs ↗69tools the agent can call
No GPUneeded to run the tests
0npm dependencies
MITopen source, every line readable
More AI work
Running in the
real world.
Two more AI products you can try or read for yourself. Like everything on this page, I built them with AI coding agents, as the engineer in charge.
Live web app · Cloud AI agent
Family Watchlist
A shared TV and movie list for households. Its assistant can search titles, add and update them and change the layout, in plain words.
- A tool-calling agent: 12 tools, up to 8 rounds per request
- Works with six AI providers; an admin picks the provider and model
- Every AI request is metered into credits and reported to Stripe usage billing, with a retry for any report that fails
- Households, email invites and Google sign-in; tested with Vitest and Playwright
Desktop app · Local or cloud models
Newsletter Creator
Point it at your sources and get a newsletter draft back, ready to export as HTML or send to Mailchimp. The AI can run on your own computer.
- A local model through LM Studio, or OpenRouter, picks the relevant stories and drafts the edition as structured data
- Every citation must point to a known source, and is checked again after drafting
- Prompts tell the model to ignore instructions hidden in source text
- Pictures from ComfyUI or fal.ai; released for Linux, Windows and macOS
Also built.
Two games, with the source on GitHub.
Dino Pals
A 3D dinosaur island for a three-year-old on an iPad. A headless Blender pipeline rigs and animates the models; three.js runs them in the browser.
View project ↗
Battle of the Old West
A card battler in Unreal Engine 5 with art prepared in Blender, a portable C++ rules engine and multiplayer that hides each player's cards.
View project ↗Working together
Prove the hard part first.
Most AI projects have one piece that decides whether the rest is worth building. We start there.
01 /
Tell me what you want built.
The task, the systems it touches and where the data can go. I'll tell you plainly whether AI is a good fit.
02 /
Build the riskiest piece.
A short paid build of the part most likely to fail, tested on your data or a safe sample of it. You see it work before committing to more.
03 /
Keep building.
We continue on an ongoing basis, with me leading the work and the team supporting it.
VINCENT SERPICOPHOENIX, AZ
Who you'll work with
I lead the work, and I write the code.
I taught myself to code at 14. I started Founders Workshop, grew it into a 30-person development company and sold it in 2026. For 16 years we built software for small and mid-sized businesses.
Every app on this page was built with AI coding agents, with me directing the work and reviewing and testing every change as the software engineer. It's how I build now, and I stay hands-on in every project.
Michael Herman
President
Michael has 30 years in solving digital problems for large brands, non-profits, startups, and small businesses, and has mentored startups since 2015.
Andrea Ramirez
VP of Operations
Andrea leads operations, process and coordination at Serpico AI. She brings more than 10 years of experience in operations and people management.
The team
Brought into a project when it needs their industry or specialty.
Daniel Love
Lynne Gabrielson
Ian Gatzke
Alexander Foster
Sabrina Valdez
David Rice
Lakshya Gupta
Kathleen Celmins
Alex Koupal
Practical questions
Before we talk.
Can it run on our own computers?
Yes. Prompt Maker and Newsletter Creator run their models on the user's own machine through LM Studio and ComfyUI. When a cloud model fits your rules, that works too: Family Watchlist supports six AI providers.
Will it work with the systems we already have?
That's the usual job. Most of the work is connecting a model to the data, tools and software you already use, and making its output something code can check.
How do we know it works?
Checks in code wherever the model's output matters, tests that run against scripted models, and evals that measure models on real requests. You see the results.
Who actually does the work?
I do, and I lead it. Michael and Andrea keep the company running, and the team brings expertise in specific industries and kinds of work when a project needs it.
Start a project
What do you want built?
Tell me about the task, the systems it touches and where the data can go. Your message comes straight to me.
Prefer LinkedIn? Message me there ↗