ExeFast

An AI economy for everyone — a public benefit purpose, chartered in FL. AI tools for underserved people, to create and strive. Self-funded or sponsored. Built and run by AI agents on Google Cloud.

Comment

Inspiration

Three fundamental principles we start from:

AI adoption is inevitable. The benefits compound, every industry is already moving, and the only open question is whether AI arrives to work for you or against you.

Inequality is a distribution problem. Abundance does not spread on its own. Systems fail at distribution and leave people behind. Agency and dignity do not come from waiting to receive a share — they come from building your own.

You know your community best. No company understands the specific needs of your street. You already know what the solution looks like. What has been missing are the tools and the skills to turn that local knowledge into something working. That is the gap we close.

The person we built for is not an early adopter with a developer on call. They have real work to execute and no team to execute it. So we made a decision that shaped everything after: if that is who we serve, the mission cannot be a marketing page we could quietly drop when it became inconvenient.

On 9 July 2026 we filed it. Article III of our Articles of Incorporation — EXEFAST INC, a Florida benefit corporation, document number P26000035240:

"Expanding access to AI tools and promoting economic inclusion for underserved individuals and communities."

Our directors are legally bound to consider it. We wrote the mission somewhere we cannot edit it when it becomes inconvenient.

What it does

ExeFast is a personal AI operator. You describe work in plain language and agents execute it — build the thing, deploy it, run it.

It meets people where they already are: the web, WhatsApp, Telegram, and Slack, with one account and one continuous thread no matter where the conversation started. The free tier is served by Gemini on Vertex AI. The platform runs on Google Cloud Run.

A voice agent your customers can simply call. Add the Voice receptionist and your business has a phone line answered by an agent that knows your business and speaks your customer's language. For an owner who could never afford someone to answer the phone, that is the difference between missing customers and keeping them. You can also customize voice agents to resell them, just by chatting on WhatsApp or other channels.

What you build belongs to you, and so does what it earns. Deploy it anywhere, charge your own customers, keep one hundred percent of what they pay you. Refer someone and you earn a flat fifty percent of what they pay us, for as long as they stay. We are not assembling an audience to monetize; we are trying to produce operators who earn, and half of what they bring us is the honest price of that.

Offline is a first-class target. Our smallest model is designed to run on-device via LiteRT-LM. A distribution problem is not solved by a product that assumes reliable connectivity.

And anyone can sponsor a seat for someone who cannot afford one — naming a recipient, or releasing seats into a public pool. Each claimed seat is a documented instance of access moving from someone who has it to someone who does not.

How we built it

The company is operated by AI agents under a written chain of command. An AI Chief Operating Officer sets direction, screens work and issues numbered rulings. An AI engineer executes them autonomously — code, database migrations, Terraform, merges to main, production rolls. Gemini handles legal and operations research. The human founder does three things: he decides, he supplies credentials the agents may not hold, and he is the final gate on anything that makes a promise to a customer.

Since 14 June 2026: 410 numbered decisions, 315 run-logs, and 29 pull requests landed on production in the final week alone. Most of that record was written by agents about their own work. The founder wrote no code.

Agents move faster than any human review can follow, so the interesting engineering was not the features — it was the machinery that makes that speed safe. Every user-facing sentence is derived mechanically from source, screened against the code that would make it true, and recorded in a claims ledger; unscreened copy blocks the deploy. Every channel ships dark and arms in two stages, the flag declared in Terraform rather than clicked into a console, with a plan-time precondition that refuses to enable a channel whose credentials are not mounted. Darkness is proven by a route being unregistered and checked against a control that could have failed — never by silence in a log.

Challenges we ran into

Agents write confident copy. At least sixteen times, published wording outran the mechanism behind it: text saying we email invitation links when no email rail existed; a "3-day Pro trial" on WhatsApp that granted no trial and no credit; plan access quietly granting a channel our pricing page sold one tier higher. Each was caught by the claims gate rather than by someone reading carefully.

Our own transparency promises broke twice. A deletion guarantee covered account data but missed conversation content stored per channel. The tripwire we wrote to catch a hypothetical future column caught a present one on its first run.

And the money path fought us for three days. Seat entitlement had never once worked — the subscription call was malformed, and a reused idempotency key replayed the cached error so three consecutive fixes were never actually evaluated. Email verification had never worked on any deploy since the first commit. Then, hours before this submission, a sponsorship bought with a real card attempted to charge twice: once for the term, once again when the seat was claimed, against a page promising a single charge. We found it by walking the product ourselves, and stopped selling before fixing.

Accomplishments that we're proud of

A working multi-channel platform, built in nine weeks by one person and a team of agents. Web, WhatsApp, Telegram and Slack; agent creation and execution; metered inference on a credit ledger; live payments on Stripe with real charges settling; legal instruments, reviewed 6 times, under a register that cannot silently drift; infrastructure fully declared in Terraform on Google Cloud, with no hand-provisioned production secrets.

A voice receptionist that answers real calls. Armed and purchasable on the serving revision, with inbound calls answered end to end in production — audio measured in both directions. Demo: 812-393-3278.

A company that catches itself. Sixteen overclaims stopped before a customer saw them. A pricing defect found by our own walk and taken offline within the hour rather than sold into. An engineer that disproved its own COO's hypothesis in writing, then refused to certify a fix it could not evidence. Our published labels currently understate what we can do — the pricing page still says Voice is coming — because a label moves only when its claim has been screened.

A mission that is a legal instrument rather than a page, filed, numbered, and binding on our directors.

And an operating record — 410 decisions, 315 run-logs — that shows exactly how an AI-run company makes, executes and corrects its own decisions. That record is what we would hand anyone asking whether this model actually works.

What we learned

Silence is not evidence. Our engineer reasoned for six hours from empty log queries while the answer sat in a field it was not reading. It ran a control, saw the control come back empty too, and kept going. A check that returns nothing when the instrument is off is not a check.

Green tests can prove nothing. Ours passed throughout, because the injected client returned success for any request — proving the caller agreed with itself, and nothing about the vendor.

Verify against what is serving. Thirty pull requests reported merged; twenty-nine reached main. One was merged into another branch twenty-two seconds after that branch merged, while every signal on GitHub stayed green. We report twenty-nine.

Nobody had walked the product. Every expensive defect above was reachable by one person clicking through the entire path once. We had automated proof of a thousand things and no proof of the one thing. Every money surface is now walked end to end before it goes public.

What's next for ExeFast

The honest starting line. Arms-length third-party revenue during the window: $0.00. Related-party revenue: $60.00 — every charge to date traces to the founder or an internal test address, reported separately as the rules require. Marketing and customer acquisition spend: $0.00. Nothing has been spent to acquire anyone. Of roughly 1,650 published strings, 807 have been read and cleared and 829 predate the claims ledger; we know exactly which.

Sponsorship reopens the moment its billing is walked clean end to end. The fix is shipped and unverified, and we do not sell what we have not proven. The rest of the platform is live and serving.

Then: the public pool page, so a stranger can see the seats waiting for them. Thunes rails alongside Stripe, because inclusion at scale means paying out where Stripe does not reach. Payouts for referrers earning their fifty percent. And the handoff loop — a user builds an agent for someone in their own community, hands it over, and keeps everything they charge for it. That loop is the mission and the business model in one motion: every handoff is a person creating economic value with AI rather than receiving it.

We are early, our revenue is nil, and our paper trail would embarrass companies ten times our size. It is also the only asset here that could not have been manufactured. We would rather be judged on it.

Built With

Share this project:

Updates