Systems we’ve already built — and what they prove.
Before we build for clients, we build for ourselves. These are products we designed and shipped end to end: the problem each one solves, what we built, where AI does the work, and the stack underneath. The same approach goes into the systems we build for you.
Photograph a thrift find, get an AI resale value backed by recent sold comparables, then turn it into marketplace-ready listings from one inventory record.
The problem
Resellers guess at value on the shelf, rewrite the same listing for every marketplace, and risk selling one item twice.
What we built
Multi-photo AI identification and condition-aware valuation, plus barcode scanning
Live web research on recent sold comparables, with sell-through and days-to-sell shown only when the evidence supports them
A buy-or-pass sourcing decision and profit scorecard
AI-written listings optimized for each marketplace — its own title, description, price, and structured fields, all editable
One inventory record across 11 marketplaces: a direct eBay integration (beta), a browser helper that fills in Facebook Marketplace, Poshmark, Mercari, and Etsy listings, and copy-ready listing kits for six more
Recording a sale lowers the quantity and queues delisting everywhere else in one step
Where AI does the work
Vision AI identifies the item and grades condition; an AI research step pulls recent sold comparables from the web and returns structured, evidence-backed pricing; then AI writes a listing optimized for each marketplace — the title, description, price, and fields that marketplace rewards. A second AI provider stands by as a fallback, and an image AI cleans up listing photos.
Built with
OpenAI
Anthropic Claude
Photoroom
Next.js
Clerk
Stripe
Postgres
Vercel
Chrome extension
What it proves: AI that turns messy real-world input — photos — into structured records your team can act on.
The operating system for regional delivery companies — order entry, dispatch, proof of delivery, billing, payments, and driver pay in one connected record.
The problem
Completed deliveries fall through the gaps between dispatch, billing, and driver pay — and work that never gets invoiced is revenue that quietly disappears.
What we built
Order intake, both manual and through a customer portal, plus a live dispatch board
A driver phone app with signature and photo proof of delivery, and a location check that flags arrivals logged far from the address
Public tracking pages for customers
Configurable pricing: accessorials, waiting time, after-hours, fuel, and customer-specific rates
Delivery-linked invoicing on each customer’s billing cycle, and driver settlements across six pay types
A revenue-leakage audit that flags completed deliveries with no invoice line, for human review
Where AI does the work
Deliberately rules, not AI, where money is decided: billing exceptions are caught by deterministic checks, each with a plain-English “why was this flagged?” explanation. We use AI where judgment helps — not where a rule is more reliable.
Built with
Next.js
TypeScript
Supabase
Stripe Connect
Google Maps
QuickBooks Online
Resend
Vercel
What it proves: One connected record from first order to final payment, so finished work always becomes revenue.
AI staff, organized: people and AI specialists working together in group chats to get real work done — with a person assigned ultimate responsibility for every task and project.
The problem
AI help is scattered across a dozen private chat windows — cut off from the rest of the team, with no shared context, no visibility, and no one clearly accountable for the result.
What we built
Group chats where people and AI specialists collaborate on real work, sharing the same context
Assign a person ultimate responsibility for any task or project — AI assists, a human owns the outcome
A Chief of Staff agent that asks clarifying questions, then coordinates seventeen specialist roles you can switch on or off
Projects with saved knowledge, so context carries over
Gmail review with drafted replies, plus calendar help — nothing sends without your explicit approval
Voice dictation and scheduled jobs
Activity, usage, and spending controls
Where AI does the work
AI specialists work alongside people in the same group chats, drafting, researching, and following up — while a named person owns each task and project. Routine and sensitive work are routed to different AI models, requests are sent with provider-side storage turned off, and a person approves every outbound action.
Built with
OpenAI
Next.js
Clerk
Supabase (row-level security)
Gmail & Calendar APIs
Vercel
What it proves: People and AI working as one team, with a human accountable for every outcome — the model we bring to client teams.
An AI strength coach that builds today’s workout from your recovery, injuries, equipment, and recent training.
The problem
Fixed programs ignore sleep, pain, equipment, and what you did yesterday.
What we built
A daily recovery check that turns sleep, readiness, and how you feel into a green, yellow, or red score with safety flags
An AI daily workout that respects injuries, available time, and equipment, with variety based on recent sessions
A workout player with form cues, reference photos, and pre-filled set logging
An in-workout chat coach and an equipment scanner that works from photos
History, a weekly recap email, and an installable phone app
Where AI does the work
Guardrails first: readiness is scored by rules, not AI, and the AI can only choose exercises by ID from a verified library — so it can’t invent one. AI output is validated before you ever see it.
Built with
Anthropic Claude
Next.js
TypeScript
Supabase
Resend
Vercel
What it proves: How we keep AI safe in high-stakes decisions: rules set the limits, AI works inside them.
Private AI Second Brain
Private — in daily use
A private AI command center that runs our founder’s day — calendar, inboxes, voice notes, tasks, and projects — from one dashboard and assistant.
The problem
Context scattered across mailboxes, calendars, a voice recorder, and many projects, with no time to pull it together.
What we built
A “start my day” briefing
Inbox triage into Needs reply, Waiting, and FYI, with AI-drafted replies that only send after explicit confirmation
Voice notes captured into a journal, with commitments turned into reminders
A multi-calendar view where the assistant proposes events instead of booking them on its own
A project board with AI-suggested next steps that cite their evidence
Secure remote access from a phone, behind multi-factor authentication
Where AI does the work
An AI assistant uses tools on your behalf, and anything risky requires confirmation. Dictation is transcribed on the device itself, so only text ever leaves the machine.
Built with
Anthropic Claude
Python
SQLite
Gmail & Calendar APIs
On-device speech-to-text
Cloudflare Access
What it proves: A true second brain — the same kind of AI assistant we build for owners who are buried in email and scheduling.
Want a system like these for your business?
Start with a free Business Systems Mapping Session. We’ll map how work moves today and show you what we’d build first.