Aliso Viejo, California

The software your business actually needs — too custom for a SaaS, too fast for an agency.

AI made building software fast. It did not make it safe. The hard part was never the demo — it's everything between "works on my laptop" and "my staff and customers depend on this daily." That gap is the whole job, and it's what I do. Working prototype in days, production in weeks, and you own every line of it.

Book a 20-minute call See the work

The problem

You've outgrown the spreadsheet, and every real option is wrong.

A core operation runs on a sprawling spreadsheet, a chain of emails, and one person's memory. Double entry. Version conflicts. No audit trail. Mistakes that cost real money. And if that person leaves, the business stalls.

Hire a developer? $120k a year and three to six months to land one — for a single internal tool.
Hire an agency? $80k and four to six months, then something generic and a maintenance headache handed back.
Buy the SaaS? Built for someone else's workflow. You pay per seat forever to fight software that's 60% right.

What you're actually paying for

"AI can build an app in an afternoon, but getting it safely into other people's hands is a whole other challenge."

Generating a working app is commoditized — anyone can prototype one in an afternoon. The billable value is the layer after the demo works: identity, access control, secret management, authoritative data connections, and a deploy path that cannot make an unsafe choice.

Every project I ship inherits the same tested security baseline, so that rigor compounds instead of being reinvented badly each time.

Services

Three ways to work together

Custom business software

Take the one workflow costing you most — dispatch, intake, quoting, scheduling, reporting, onboarding — and turn it into real software shaped to how you actually work. Connected to your real data, with proper accounts and permissions from day one.

AI-powered products

Real-time voice and video, automated grading and analysis, agents that act on your systems through typed, audited interfaces. Built by someone who has shipped all three — and who knows where AI features quietly fail once real users arrive.

AI-managed websites

A custom site you never log into. Text or email the change in plain English and it's done, with a preview before anything goes live. Every change versioned and instantly reversible. No CMS to learn, no plugin updates.

Selected work

Shipped, not slideware.

Systems running in production, not mockups. Each is labelled honestly — client engagements, my own products, and internal tools are different kinds of evidence, and you deserve to know which is which.

Kranson-Schubert Mental Health Foundation

Client · Live

Dr. Donald K. Schubert, a psychologist of 46 years, was launching three things at once: a mental-health foundation named for his parents, a first-annual community Thanksgiving celebration, and an author catalogue for his five books. He had no web presence — he'd tried a DIY site builder a year or two earlier and abandoned it as too confusing.

I built three cross-linked sites with a nine-theme design system he chose from, a password-gated editor, a Google-Doc editing pipeline with fuzzy matching so a typo corrects a record instead of duplicating it, and plain-English editing by text message. Live on custom domains eleven days after the first call — with the first deploy the same day we spoke. The domains were registered in his name, not mine, so the Foundation owns its identity outright.

kransonschubertfoundation.org missionviejothanksgiving.org drdonaldschubert.com

Static HTML/CSS/JS, no framework · content-file architecture · Vercel · GitHub Actions CI/CD · git-backed headless CMS committing through the GitHub API — ~17 pages across 3 sites

Ariya MD

Own product · Live

A platform where medical students and international graduates book peer-to-peer video mock residency interviews and medical-English speaking practice. Sessions are recorded and transcribed, then calibrated AI graders return per-criterion scored feedback. Availability matching, scheduling, forums, admin analytics, and subscription billing.

ariyamd.com

React 19 · Firebase · Vercel functions with bearer-verified auth on every route · Daily.co · Gemini · Stripe — token-verified API surface, audited admin impersonation, grader-drift regression harness

Construction OS

Prototype

A pre-construction pipeline for contractors: lead → plan documents → material takeoff → estimate → proposal → e-sign. The architectural point is that every mutation flows through one typed tool registry shared by the UI, an in-app agent, and outside AI over an MCP endpoint — so each change is automatically audited and attributed to a person, an agent, or an external actor. Built with a working contractor as domain expert and acceptance tester.

Next.js · TypeScript · zod · zero extra npm dependencies, including a hand-rolled MCP server · pluggable store — 25 typed tools across a 38-module registry

Kith

Own product · Beta

Video calls where each person picks one language, speaks it, and hears everyone else in it. The organizer has an account; a relative just taps a link. A production translation agent runs one live session per speaker-to-language pair, publishing translated audio and captions per listener, and skipping same-language pairs so they cost nothing.

Next.js 16 · Firebase magic-link auth · LiveKit Cloud agent · Gemini Live Translate · dual-clock scheduling with property-tested timezone handling

PhoneCam

Tool · Built

Turns an Android phone into a 1080p30 wireless webcam. The load-bearing requirement was that streaming survive screen-off and app-switching — which every browser-based approach fails — solved with a camera-type foreground service streaming hardware-encoded H.264 over WebRTC. Lens switching from the notification, live stats overlay, thermal auto-downgrade, and camera-steal detection with auto-resume.

Kotlin · Camera2 · WebRTC/WHIP with shared EGL · MediaMTX · OBS — replaced a commercial app after an instrumented network post-mortem showed why it failed

Book Reader

Tool · Daily use

A mobile-first reading app for long-form markdown: built-in library, uploads, live-from-disk files, and article URLs extracted server-side into clean reading copy. Chapter navigation with full-book search and resume-at-last-sentence. Two features earn their keep — highlight-to-note with typed categories that compile into a machine-fixable feedback report, and sentence-level narration with tap-a-sentence-to-jump using local voices.

Python stdlib server, no framework · vanilla JS PWA · Kotlin WebView wrapper for Android · local neural TTS — runs as a service on both desktop and phone

Policy simulation model

Own work · Finished

A single self-contained HTML file modelling a contested public-policy question. Drag sliders for the disputed assumptions and two independent estimators re-run live, with widely-cited public figures marked on the resulting distribution. The estimators are deliberately kept side by side rather than averaged, so the difference between counting something and forecasting it stays visible to the reader.

470 lines, zero dependencies, zero build, any static host · ships with a full parameter-to-source provenance table

AI grader factory

Internal · Tested

The quality apparatus behind automated grading: a verbatim-overlap similarity guard, a sufficiency gate that refuses to grade an under-length session rather than inventing a score, a deterministic worksheet builder, a grader-agnostic calibration harness, and a drift sentinel that re-runs production graders daily and control-charts band stability. The unglamorous work that keeps an AI feature trustworthy months after launch.

Node ESM · node:test · systemd · strict provenance headers pinned to an upstream commit — 105 tests passing

Happy to walk through any of these live — including the code and the reasoning behind the decisions. Other client work is covered by confidentiality and shown only on request.

Process

How a project actually runs

Watch the work

I sit with the process as it happens today. The real problem is usually not the one described on the call.

Prototype in days

Something running that you can click, not a document. We find out fast whether the shape is right.

Harden it

Accounts, permissions, secrets, real data connections, and a deploy path that can't ship something unsafe.

Hand you the keys

Your repo, your cloud accounts, your domains. Keep me on retainer or don't — either way it's yours.

Fit

Let's find out fast whether this is a match.

A bad-fit project costs us both more than it's worth. This works well when:

There's one specific workflow costing you real time or money right now.
You have real data the software needs to be built around.
The person I'm talking to feels the pain and can decide — no procurement committee.
You'd rather own a safe custom system than buy the cheapest possible build.

Probably not a fit

Shopping purely on hourly rate.
Pre-revenue idea stage, looking for an equity co-founder.
No existing process and no data to build around.
Research, hardware, or custom ML training projects.
Inheriting an undocumented mess sight unseen, fixed price.

If that's you, say so on the call and I'll point you somewhere better. Saying no early is a feature.

Tell me what's broken.

Twenty minutes. Describe the workflow that hurts most and I'll tell you straight whether software fixes it, roughly what it would take and cost, or that you don't need me at all.

hello@jzmlabs.com