★ Six builds · six kinds of AI no black box Every tool named · nothing invented Real, prototype or concept, kept straight American-built · New Hampshire ★
Case studies · six builds, six kinds of AI

What actually
got built.

Six builds, six different kinds of AI. Each card shows what was broken, what got built, every tool used, and what's real versus what's still concept. That last part is on purpose.

the problem → the big fix → the workflow · every tool named · nothing invented
01 The morning brief 02 Claims that file themselves 03 Team celebrations 04 Tribute films 05 A self-updating site 06 An app from one idea
01the origin
01 · Worked example · the origin

The HVAC Shop

a one-truck shop · the first build

The problem

A one-truck shop runs out of one overloaded head. Which filter fits this unit, how much refrigerant that system holds, the gate code, what's due this week, it all lives in a spreadsheet nobody opens on the way out the door.

▼ so I built

The big fix

A pre-job brief, written from the shop's own equipment sheet: the filter first, the refrigerant amount, with the EPA tracking flag only when it crosses the 15-lb line, the access notes, the due date. Before every job, plus a morning rundown of the week.

▼ here's the workflow
What comes in
The equipment sheet
Google Sheets · one row per client
What happens
Mr. Wilson
  • Zapier posts the job's row
  • he replies with the brief in plain English
  • a scheduled Zap writes the morning rundown, soonest job first
What you get
A brief before every job
plus the morning rundown of the week

Workflow type: no-code automation, spreadsheet in, plain English out.

Built with:

Built overnight against a fictional demo shop, the origin prototype every later trade copies, not a live client.

"Mr. Wilson started as a spreadsheet, a Zapier flow, and one rule: tell the tech what they need before the job, and never make up what you don't know."

02client blinded
02 · Vision Model · Image Detection

Regional Insurer

a regional carrier · client blinded

The problem

Every claim that hit the mailbox got opened, read, and re-keyed by hand before a decision could even start. Fraud signals, doctored receipts, edited photos, were caught by eye, if at all.

▼ so I built

The big fix

A pipeline that clears the claims mailbox: pulls the claim number, screens the attachments, routes by risk, and files everything into the carrier's claims system. People only touch what needs judgment.

▼ here's the workflow
What comes in
A claim email + attachments
Outlook · Power Automate
What happens
The claims pipeline
  • extract the claim number, pattern-match first, AI only when unsure
  • screen attachments for fraud signals
  • route by risk
AI fraud-risk score concept
today's screening is rule-based, and we say so
What you get, routed by risk
low riskFiled automatically
everything elseFlagged for a person

Workflow type: document AI, reading and filing, not chatting.

Built with:

A proof of concept, tested against mock data, not yet wired to the live carrier.

"The fastest AI win in most businesses isn't a chatbot, it's deleting the paperwork nobody should be doing by hand."

03running workflow
03 · Node-based AI

Dental Office

employee recognition, built inside a dental-tech company; their product isn't mine

The problem

The team's daily experience of AI was headlines and hype. Recognition, meanwhile, took somebody hours per person, so mostly it didn't happen.

▼ so I built

The big fix

A no-code node workflow: one photo and a one-line note in, a personalized celebration film styled to that person's favorite show, plus a branded award, posted straight to the team's channel.

▼ here's the workflow
What comes in
A photo + a one-line note
dropped into the Krea canvas
What happens
A Krea node graph
five models, each one job, wired in order
  • read the note, pick the theme
  • cast the person into the scene
  • lock a real film look
  • animate and score it, then build the certificate and post
What you get, in #celebrations
A celebration film
styled to their favorite show
A branded award
matching certificate

Workflow type: node-based, multi-model, five AIs on one visual canvas, no code.

Built with:

Real and running, week after week. No engagement numbers quoted, the honest proof is people asking "who's making these?"

"The best way to get a team excited about AI isn't a training deck. It's making them the star of something delightful."

04real engine · in pilots
04 · Multi-model AI

Funeral Home

ruby-v1 · a tribute-film engine · client blinded

The problem

A grieving family has a shoebox of photos, some home video, and a song, and nothing between a PowerPoint slideshow and hiring a creative agency. When my own mother passed, that gap was personal.

▼ so I built

The big fix

An engine, ruby-v1, named for my grandmother, that reads everything the family brings, derives one custom look for that one person, and renders a beat-synced tribute film plus the matching keepsakes: prayer cards, programs, signage.

▼ here's the workflow
What comes in
Photos
Gemini vision
+
Home video
Twelve Labs
+
The song
librosa
+
The obituary
Claude
four kinds of memories → one engine
What happens
ruby-v1 · the engine
Before a frame is rendered, the engine settles one question: what should this person's film look like? Nothing comes off a template shelf.
  • derive one look for this one person, then lock it
  • caption every photo
  • mine the home video for the moments that matter
  • cut to the song's beats, in the locked era grade
Inside that first step · the identity lock
Claude writes the prompt
from the obituary and the family's answers
free rein on style, one required object: a rocking chair
Midjourney paints it
one image, made for one family
not a preset, not a filter
The family approves it
they pick the one that looks like the person they lost
the one human gate in the middle
Gemini reads it back
palette, typeface, era grade
written down as the family's design system

plain englishThe approved painting carries a style code. Every image the engine makes after that point carries the same code, so the film, the monogram and the printed keepsakes all come out looking like one family and no other. The look is derived, not chosen off a menu, which is why no two families get the same film.

What you get
A tribute film
scored to their song
A matching keepsake set
cards, programs, signage

Workflow type: multimodal pipeline, six systems each doing what it's best at, with ffmpeg as the render step.

Built with:

The engine is real and has made tribute films for real families, mine first. Pre-revenue, in pilots with funeral homes; I'd rather say that plainly than invent traction.

"The point of AI here isn't speed. It's care at scale, the kind of memorial that used to take a creative agency."

05working prototype
05 · Agentic AI

Community Church

a self-updating church site · client blinded

The problem

A 125-year-old church run by volunteers. Every Friday someone re-typed the bulletin into the website by hand. At-home viewers couldn't follow the song lyrics on the stream. And 125 years of records sat fading in a basement.

▼ so I built

The big fix

One five-minute upload: drop in Friday's bulletin PDF and the site refreshes itself, services, songs, scripture, a fresh header image. A live-lyrics overlay for Sunday's stream and a searchable 125-year history portal round it out.

▼ here's the workflow
What comes in
Friday's bulletin PDF
one five-minute upload
What happens
The Friday refresh
  • AI reads it into clean data
  • the site republishes itself
  • the lyric overlay arms from the songbook files the team already keeps
What you get
The site, current for Sunday
republished automatically
Lyrics on the stream partial
built, still running demo lyrics

Workflow type: agentic content system, one document drives everything downstream.

Built with:

The Friday upload works end to end. The live lyric sync is built but partial, still running demo lyrics. The history portal (75 searchable events) is the most finished piece.

"AI doesn't have to replace the people who make a place special. Sometimes it just gives them their Friday back."

06vibecoded prototype
06 · Vibecoded apps

Youth Sports

a coaching tool, built solo with AI

The problem

A volunteer coach runs a whole team out of a roster spreadsheet, parent group texts, and a game-plan notebook. A real tool to replace that pile used to take a development team and a budget.

▼ so I built

The big fix

A complete coaching tool, brand, feature set, product spec, clickable prototype, built solo with AI, no code written by hand. Underneath it, a real algorithm that generates fair lineups and playing time.

▼ here's the workflow
What comes in
One coach's idea of the tool he wanted
described in plain English
What happens
The vibecoding loop
the only place the loop appears on this card
1 Describe it
→
2 AI builds it
→
3 Refine
The AI lineup engine concept
a real deterministic algorithm exists, not yet wired into the UI; the in-demo "AI game plan" is a stand-in
What you get
A clickable coaching prototype
roster, schedule, comms, playing-time planner

Workflow type: vibecoding, a 0→1 product, spoken into existence.

Built with:

A designed, working prototype, no users yet.

"One person described the tool he needed, and it exists. That's the whole reason a done-for-you service can exist for the shop down the street."

the trades he knows

Every trade has its own busywork.

The tracking, the follow-ups, the deadlines nobody has time for. He shows up already knowing yours.

HVAC

Job briefs, refrigerant logs, and deadlines that sneak up.

Auto Repair

Recall lookups, customer updates, audit-ready records.

Salon

Appointment-linked inventory, reminders, safety sheets.

Vet

Reminders, records, controlled-substance logs.

Daycare

Parent updates, staffing ratios, licensing dates.

Solar

Permitting timelines, customer questions, incentive deadlines.

Home Inspection

Scheduling, fast reports, certificates on time.

Insurance

COIs, renewals, and claims paperwork, sorted.

Property Mgmt

Tenant follow-ups, filing dates, certificates on call.

CPA / Accounting

Doc chases, filing deadlines, the crunch smoothed.

Self-Storage

Lien notices and state deadlines, never missed.

Your trade

Whatever eats your week. If it repeats, he can take it.
Same idea, your shop.

Want one of these for your trade?

Start with the audit. No card, and you keep the brief either way. The first fix is built and installed before you commit to anything ongoing.

Start with the audit no card · keep the brief