Bespoke
We build apps around the way your company already works and run AI assistants on machines in your office. Your bids, drawings and customer files are never sent to an AI company.
We build AI assistants around the way your firm already works, connect them to your own software, and run them on machines in your office. The assistants never send client files to an AI company.
We rebuild outdated websites for local businesses and build AI assistants that take repeat paperwork off your staff. Your website is hosted online, like any website; the assistants run on machines in your office and never send customer files to an AI company.
Every project starts with a free review: a 20-minute call or visit, then a one-page map two days later. No charge, nothing to sign.
Three ways in · Remote or in person
The free review is a 20-minute call or visit. Two days later you get a one-page map: the jobs worth building software for or automating, the hours each takes now, and what each would take to fix.
For construction, fabrication and the trades
Bespoke apps for estimating, the shop floor and the field, and assistants on your own machines for bids, quotes and paperwork.
Where we usually start
A quoting tool · a daily report app
For professional firms
An assistant on your own machines, connected to your practice software, for law firms, clinics, accountants and advisers whose rules make cloud AI awkward.
Where we usually start
Intake summaries · first drafts · search across past work
For local businesses
A website of up to five pages on your own domain, built for phones, with one AI feature. We also build assistants that take repeat paperwork off your staff: quote replies, order confirmations, estimate follow-ups.
Where we usually start
A website rebuild · replies to quote requests
Moonlit Lab AI is a studio in the Northeast USA. We work with businesses anywhere, by video call and remote access you approve, or in person. Your website is hosted online, like any website; anything that reads customer files runs on your own machines.
Bespoke software
Off-the-shelf software makes you work its way. We build small, exact tools around the way your people already do the job, for a Mac in the office, an iPad on the shop floor or an iPhone on site. Each one starts from a job we can time.
Your own price book, labor rates and markups in one tool, so a takeoff or a customer's RFQ becomes a quote the same way every time.
For estimators and the office.
Job travelers, daily reports, punch lists and time entry on the iPad or iPhone your crew already carries.
For the floor and the site.
Jobs, hours and margins on one screen, pulled from the accounting, scheduling and estimating software you already pay for.
For whoever runs the jobs.
The workbook or old database the business runs on, rebuilt as an app that checks what goes in and keeps a history of every change.
For the person who keeps it alive today.
We build our own software. Moonlet, our on-device AI app for iPhone and Mac, and the engine that runs it are written by us. Moonlet is in development.
Apps we build
Built for iPhone, iPad and Mac around your jobs, your names and your brand. Apps for your team go straight onto your company's devices. Apps for your customers go on the App Store under your company's name, published from your own Apple developer account, so the listing and the customers are yours.
Tuesday, 21 October
Clock in
On site since 7:02 AM
Crews clock in and out against a job and a task code, breaks included, and the office gets timesheets ready for payroll. It keeps working without signal on site and catches up when it reconnects.
For crews, shops and field teams
Warehouse A
Stock
Scan a barcode to receive, move or use stock, see what is running low before a job needs it, and count a shelf in minutes instead of an afternoon.
For the warehouse, the shop and the van
Marketing
Campaigns
Send a seasonal offer or a review request after each job to the customers who said yes to hearing from you, and see which ones came back.
For the owner and the office
Your next visit
Thu 23 Oct, 9–11 AM
Recent
Your customers book a visit, follow the job, approve a quote, pay an invoice and message your office, in an app with your name on it in the App Store.
For the customers you send to it
AlsoSafety checklists and inspections · equipment and maintenance logs · delivery and dispatch · quote builders · customer portals
Sample screens with made-up data. Apple reviews every App Store app; a real app with its own features passes, a website in a wrapper usually does not, and we build the first kind.
Your client files stay in the office. The AI comes to them.
Why it stays in the office
Every field here answers to someone for its files: defense shops answer to DFARS and CMMC, contractors to their clients' contracts and NDAs, and law firms, clinics, accountants and advisers to strict confidentiality rules. An assistant on your own machines takes the AI company out of that question.
General contractors, builders and remodelers
Contracts and NDAs, not one statute
Most construction work answers to contracts rather than a privacy law: owners' drawings under confidentiality clauses, subcontractor prices given in confidence, and your own estimates, which are what win you work. Defense and federal projects can add rules for drawings marked CUI. Every assistant we build runs on your own machines, so none of it goes to an AI company.
Source: DFARS 252.204-7012, for drawings marked CUI on defense work · ABA Litigation Section on Opinion 512
Where we usually start
We make compliance simpler. We never call a setup compliant, because your own counsel or adviser has the final word. These summaries come from the rules' own text or professional commentary and are not legal advice.
Private AI
Every build starts from a task your people already do by hand and can time. We connect the assistant to the software where that task lives, and a person checks everything it writes.
For your construction company
A bid package or spec book, summarized on one page: scope, dates, insurance and the clauses to price.
Documents in, one clear page out, in your format.
For your construction company
RFIs, change-order letters and daily reports, drafted from field notes for a project manager to check.
Letters and replies drafted for a person to check and send.
For your construction company
Plain-English questions across past bids, specs and submittals kept on your own machines.
Questions in plain English across what your firm has already written.
For your construction company
A connector to the estimating, scheduling and accounting software you already use.
The assistant works inside the software your team already uses.
The assistant drafts and your team decides. Every build is set up so a person reviews each output before it goes out.
Websites
Outdated, insecure or broken website? We rebuild it into a fast, modern site that looks sharp on a phone and is secure from day one.
For a practice under HIPAA, a website form that collects patient details is itself an online service. Your adviser decides what the form may ask and which form service it uses, such as one that will sign a business associate agreement.
Plumbing, 7 days a week
Leaks fixed today, not next week.
Licensed plumbers across the county, with the price agreed before we start.
Services
Get a quote
Reply drafted. Your team checks it and sends it.
Made-up exampleA plumbing business's site, rebuilt: home, services, a quote form whose replies an assistant drafts, and the phone view.
The AI stack
We wrote our own engine, the software that runs the AI, and it powers Moonlet, the on-device learning app we are building. It is one tool among several. We also install and tune the open engines and agent tools teams already use, such as llama.cpp, vLLM, Ollama, MLX, OpenCode and Hermes Agent, and choose whichever runs your model best on your machine. The model, the software, the hardware and the build all come from one studio.
Agent tools we connect to your model
An agent tool has no speed of its own. It runs as fast as the engine and the model behind it.
| Our engine on a MacBook Air, M5, 16 GB | Result |
|---|---|
| Ornith 35B writing a reply | 13 tokens/sMeasured, Oct 2026 |
| First words of a reply to a short request | 2.3 sMeasured, Oct 2026 |
| While another app holds 5 GB of memory | about 1% slower12.6 → 12.5 tokens/s, Jul 2026 |
| Each extra open conversation, 65,000 tokens (about 100 pages) long | 0.3 GBplus 90 MB of model state; calculated from the cache format, not measured |
| Open engines, measured by others | One person asking |
|---|---|
| llama.cpp on an NVIDIA DGX Spark, gpt-oss-120b | 55 tokens/sNVIDIA, Oct 2025 |
| vLLM on a DGX Spark, gpt-oss-120b | 61 tokens/s116 to 153 tokens/s in total with eight people asking at once. A DGX Spark owner on NVIDIA's developer forum, Aug 2026 |
| Ollama in the same test | 42 tokens/sno faster in total with eight people asking at once; same post |
| MLX on a Mac Studio, M3 Ultra, 512 GB, Qwen3 235B | 24 tokens/s16 tokens/s on the same Mac in the GGUF format. MacStories, 2025 |
Each figure comes from its publisher's own test, so compare rows from the same source. The engine, the format and the number of people sharing a machine change the result as much as the hardware does, which is why we measure on your machine before the build starts.
Watch a three-minute demo on a MacBook with the Wi-Fi offEvery document in the video is made up.
The hardware
We size the machine from the model and how many people will use it, then install it, set it up and look after it. On Macs, the assistant runs on our own engine. On NVIDIA hardware we install proven open-source inference software, such as vLLM or llama.cpp; support for our own engine there is in development.
One desk, or a small office
Small, quiet and low on power. A Mac mini, with up to 64 GB of memory, suits one or two people; a Mac Studio, with up to 512 GB, runs the largest open models for a team.
Runs our own engine
NVIDIA on the desk
A desktop AI computer with 128 GB of memory. NVIDIA rates one for models of up to 200 billion parameters, and two linked together for up to 405 billion.
Runs open-source inference software
Fast answers for a small team
A tower built to the model and the budget, around NVIDIA GPUs such as the GeForce RTX 5090, with 32 GB, or the RTX PRO 6000 Blackwell, with 96 GB.
Runs open-source inference software
Many people at once
Rack servers with several NVIDIA GPUs, for a server room or a closet with the power and cooling for them. We size and quote them after the audit.
Runs open-source inference software
The pictures are our own illustrations, not product photographs. Memory and model-size figures are the makers' own; prices change, so every machine is quoted in writing when we size it.
The fitting
These are the five steps of a full build of a bespoke app or a private assistant, including the audit. A website rebuild or a starter automation skips the audit: the free review's map leads straight to a written quote and timeline. A website takes 2–3 weeks.
Free · 20 minutes
A 20-minute call or visit. Two days later you get a one-page map: the jobs worth building software for or automating, what to fix on your website, and the hours each takes now. No charge, nothing to sign, and the map is yours to keep.
Weeks 1–2
We time the job as it is done today, choose one or two builds with a measurable target, size the hardware and write your data-handling note.
Weeks 3–4
One app or automation for two or three staff, measured against the baseline, with a check-in each week.
Weeks 5–6
We connect it to your software, train your staff for an hour and hand over a written guide.
Every month
Model and app updates, monitoring, one new workflow each quarter and a monthly report of the hours saved.
Team training
Claude Code is Anthropic's AI coding tool: you describe what you want in plain English, and it reads files and writes and runs code on your computer. We teach your staff to use it on their own work, each on one small tool for their job, such as turning a spreadsheet into a weekly report.
On your machines
The assistants we build
Anything that reads customer files runs here, in your office, and is never sent to an AI company.
Anthropic's cloud, your own account
Claude Code training
What your team types into Claude Code, and the files it reads, are sent to Anthropic's servers.
On site, for up to 8 people. Each person works on one small tool for their own job.
Half a day
One two-hour session a week, plus office hours, working toward a tool each person's team can use.
Four weeks
A monthly session for staff who keep building, added to a care retainer.
Every month
Your data
Every private AI audit ends with a short written record for clients, partners or a regulator. It is prepared for your firm and reviewed with you, and it sets down four things about your assistant.
Data-handling note
Prepared for your construction company
Set up with FileVault disk encryption and separate user accounts, with client folders kept out of cloud sync.
Questions
Yes. It is a 20-minute call or visit about one job your people repeat every week, or your website. Two days later you get a one-page map: what is worth building or automating, the hours each takes now, and what each would take to fix. No purchase is needed and there is nothing to sign. If the numbers don't justify a project, you keep the map.
Yes. Everyone starts with the free review, and most first projects are one small tool, one automation or one website rebuild. A full private build starts with the audit. Every job is quoted in writing before any work begins.
Not for the free review. The website's AI feature and every assistant we build run on a machine in your office: a Mac, an NVIDIA DGX Spark, a custom workstation or a GPU server, sized from the model and how many people will use it. A Mac you already own may be enough. If you need a new machine, you buy it directly from its maker, and we tell you which one before work begins.
Yes. That is most of the work: a tool for one job, built around the way your people already do it, for the Mac, iPad or iPhone. It starts from a job we can time, so you can see what it changed.
Yes. It is published from your own Apple developer account, which Apple charges for yearly and checks your company for first, so the listing is in your name and the customers are yours. Apple reviews every app before it goes live: a real app with its own features passes, a website in a wrapper usually does not.
Five things: whether the address bar shows the padlock, whether the site works on a phone, whether it looks up to date, how fast it loads on a phone, measured with Google's PageSpeed Insights, and whether customers can book or ask for a quote online. We look the way any visitor would. We never submit your forms or run scanners on your site.
No. It is hosted online on your own domain, like any website, so your customers can reach it, and a message sent through its form travels over the internet like any web form. The privacy promise on this page covers the assistants: anything that reads customer files runs on your own machines.
Keep them. The assistant that takes repeat paperwork off your staff works with whatever site you have.
No setup makes a firm or business compliant on its own. Keeping client and customer files on your own machines makes the question simpler, and the data-handling note records how it works. Your counsel or adviser makes the call.
Not for the AI. The model runs on your own machine and reads files there. A connector may reach software you already use online, such as a cloud practice-management system, but the assistant never sends your files to an AI company. Your website, if we build one, is online like any website.
Only when we work on your machines. The assistant never sends files to us. Setting it up and looking after it can mean working on your machines, in person or through remote access you approve, and that access is written into the contract. Your counsel or adviser decides what agreement it needs, such as a business associate agreement under HIPAA.
No. The assistants we build run on your own machines; Claude Code training uses Anthropic's cloud under your own account. Your team uses seats on your business's Claude Team plan (Enterprise for larger firms), bought from Anthropic, not personal Pro or Max accounts, and files with customers' personal information never go into Claude Code.
It drafts and your team decides. Each build is set up so a person reviews every output before it goes out.
One Mac running a large model from its drive serves one or two people asking the assistant to write at the same moment. Larger teams get a bigger machine: a Mac Studio, a DGX Spark, a workstation or a GPU server, sized before the build.
Open models with open licenses, such as Ornith (MIT) and Qwen (Apache 2.0). You pay no per-seat AI fees for the assistants we build, and the model on your machine keeps working whatever an AI company does with its prices. Claude Code training is the exception: each trainee needs a seat on your business's Claude Team or Enterprise plan, bought from Anthropic.
Free review · 20 minutes
Tell us about one job your people repeat every week, or your website. Two days later you get a one-page map of what to build, what to automate, the hours each takes now and what each would take. No charge, nothing to sign.
Fill in the card, or email us at hello@moonlitlabai.com. Say what your business does and name a day and time that suits you, and we answer within one business day.
Worth a free review if you have