Who Built This, and What We Owe Them

Start with the number that is not what it looks like

This repository is three months old. The first commit in it is Ankit Kumar Singh's, dated May 5, 2026. Since then it has taken 1,477 commits.

Six hundred and forty-nine of those, forty-four percent, were written by an AI agent committing under an Anthropic no-reply address. Eight hundred and twenty-four were written by six people. Four came from an account we have not yet matched to a name, which is a small failure of our own record keeping and is now on somebody's list. We are opening a thank-you note with that breakdown because the sentence "our team shipped 1,477 commits" would be a lie told by omission in the first line, and nothing written after it would be worth reading.

The company is older than the repository. The brand dates to 2021, and the rebrand to 🤫 hussh relaunched on April 1 this year. Three months is the age of this codebase, not the age of the work, and we are not going to stretch it into a first anniversary.

What we set out to do

One thing, stated plainly: connect a person who has a problem to a named human nearby who can solve it. Every answer engine on earth will give that person an answer and then leave them needing an actual human being to close the loop. That human is the product. Around it sits 🤫 Private Agent One, which is free to every American citizen 🇺🇸 and always will be, because the moment a citizen is the one paying, the revenue has to come from their data or their attention instead.

The constraint that makes it hard is the part we are proudest of. Never fabricate a listing. Cite the source on every record. Let anybody leave, for free, with no account and no retention flow. A directory you can check is worth something. A directory you cannot check is worth nothing, and there is an entire industry built on the second kind.

What actually got built

In the sixty days to August 6 there were 980 first-parent commits on main. What they add up to:

• The directory went from a handful of curated postcodes to 99.82 percent of the US map at the thin tier, covering roughly 1.16 million licensed professionals, with the honest coverage number printed on the page rather than the flattering one.
• Insurance producer rosters are live for six states, Texas, Florida, Louisiana, Iowa, Ohio and Nevada: 719,265 resident producers across 6,560 postcode areas. Ohio alone was 65,872. Forty-four states still have no insurance inventory at all, and our board KPI page says so in the same size type.
• Healthcare individuals seeded from the public NPI registry. Community organisations, the food bank and the crisis line and the place that helps with bills, seeded from the IRS exempt-organisations file: 369,447 organisations across 31,964 postcode areas.
• County front doors for King County, Washington and Clark County, Nevada. Clark County is 76 postcode areas of which our roster covers 69, and the page says both numbers.
• Removal made real, in three separate commits, because it turned out that only one of three code paths that emit a person ever imported the suppression check.
• The consent layer: a published scope registry at v0.5.0 with 252 scopes, a backend grant path that now fails closed on an empty request and on an unpublished scope and writes a deny receipt either way, and, on August 6, the first surface where a person sets a real scope themselves.
• Our own applicant tracking system, built in a single day.
• The test suite went from 1,219 tests on July 28 to 2,200 tests across 228 files nine days later. The whole src and backend tree contains exactly two TODO comments.

The six people

Kushal Trivedi, co-founder and forward-deployed core engineer, 592 commits. The recent pull request stream, #440 through #515, is entirely his. The most honest way to describe his last three months is to point at the section above and say that most of it is him.

Ankit Kumar Singh, founding engineer, 184 commits, including the first one in the repository. Eleven merged pull requests, most recently the careers application page in July.

Jhumma Kumari, UI/UX lead, 23 commits under a hushh.ai address. A commit graph systematically undercounts a design lead, because the work arrives as a decision somebody else then types. Twenty-three is the wrong number for what she changed and we do not have a better one.

Akshat Kumar, AI full stack engineer, six commits and three pull requests in July: the OpenClaw Hermes marketing pages and the backend site-context routes.

Gautam Ahuja, ten commits. His pull request in July shipped the feature-by-feature Agent One marketing pages.

Neelesh Meena, nine commits across five pull requests, from the very early ones to the One Location marketing writing and imagery that landed between July 12 and August 5.

There are zero outside contributors to this repository. Every pull request in it was opened by somebody on that list. If we thanked "the community" for this codebase we would be thanking nobody, so we are not going to.

The other forty-four percent

An AI agent wrote 649 of these commits. We are not thanking a tool. We are telling you it is there, and telling you what it is actually like, because the version of this you usually read is either a marketing claim or a confession, and it is neither.

It looks like this in the log. "WIP checkpoint: consent scope validation, agent still working." "Checkpoint two agents mid-run: NY attorney pull and the certification samples." At one point, while somebody was measuring the state of the tree for an internal status document, a second agent modified two files underneath them and the measurement quietly stopped describing reality. That is a real coordination problem and it is new.

The interesting commit is this one: "Correct a false claim: hu_ssh is a design, not shipped code." Somebody, human or agent, had written a page describing a transport as though it existed. The correction is in the permanent record next to the mistake. That is the whole working arrangement in one line. Speed is cheap now and confidence is cheaper, and the only thing that keeps the pair of them from producing a beautiful codebase full of lies is a gate that runs.

Ours runs thirteen checks in order before anything merges: a phone-auth guard, secret safety, brand, emoji, copy standard, directory integrity, design bar, design system, gate parity, environment parity, types, the full test suite, and a production build. Every single one of them exists because something got past us once. The reason we can let an agent write forty-four percent of the commits is not that the agent is trustworthy. It is that nothing here is trusted, including us.

What we got wrong, including this morning

A gratitude post that lists only wins is an advertisement.

Today, in production. The directory search API was throwing away its own ranking. Every record in a postcode geocodes to the same centroid, so a final sort by distance was a sort on a total tie, which restored the source file's order and discarded the quality ranking entirely. A request for home services in one Connecticut postcode returned two limited liability companies above two named humans, alphabetically, all the way down. In the same response, every result carried a button reading "Verify at the regulator" that pointed at our own claim form, so a person trying to check somebody against a state regulator was instead asked whether they were that person. And a search for community organisations returned the same parent-teacher council three times in the top eight. All three were reproduced against the live site, and the fix was written and tested the same day.

Twice in one day, in July. A job application form told candidates it had sent applications it had not sent. It was fixed, and then the same defect was found in a second form the next day.

The same week. A support rail we had announced as working had never sent an email. Our public releases page went 59 commits stale, which means an entire day of shipped work, the applicant tracking system included, was missing from the public record while we were describing ourselves as a company that builds in the open.

Earlier. We published 47 consent scopes and the server implemented eight. The other 39 returned a clean 200 and nothing at all, silently, for an unknown length of time.

The pattern in all of it is one thing: announcing that a surface works before confirming that it works. Not carelessness. Each claim was true when somebody wrote it, or was about to be true, and nothing in the build compared the claim to the behaviour. Our answer is a rule rather than a resolution to be careful. Any statement that can be enforced must be enforced by something that runs.

While we are being precise: FedRAMP High and the DoD impact levels are in pursuit, not held. Our certifications page says so, and it will keep saying so until a third-party assessor or an agency authorisation says otherwise.

The public record, and what we actually owe it

Most of this product is made of data that American institutions publish because a law says the public should be able to see it. The relationship is worth stating exactly, because it is easy to dress up: they publish, we used what they published, and not one of them has approved of us, endorsed us, or heard of us.

The Census Bureau's 2024 gazetteer of 33,791 postcode tabulation areas is the proximity backbone of the entire directory, and its county relationship file is the reason Pahrump is not filed under Las Vegas. The Centers for Medicare and Medicaid Services publish the NPI registry that our healthcare records come from. The Internal Revenue Service publishes the exempt-organisations file behind every charity and community record, and the search tool that lets a reader check one. The SEC publishes Form ADV and IAPD. FINRA publishes BrokerCheck, the CFP Board publishes its verification tool, NIPR publishes the national producer lookup, the NAIC publishes the state department directory, and Washington's Office of the Insurance Commissioner publishes the door we link every Washington producer back to.

The state insurance departments of Texas, Florida, Iowa, Louisiana, Nevada and Ohio publish producer rosters as official open data, and we parsed them. Contact details, street addresses, emails and phone numbers, are dropped at parse time, before anything is written, in every one of those pipelines.

Iowa is different in one respect and it matters. The Iowa Insurance Division publishes its producer dataset under CC-BY-4.0. Attribution there is not a courtesy we are extending in a blog post. It is a term of the licence we are operating under, and we would owe it to them even if we felt nothing.

Two more, for something we did not build. The Free Law Project's CourtListener bulk release and the Harvard Caselaw Access Project are genuinely excellent public infrastructure, cleanly licensed, with nobody to negotiate with. We measured 166,806 rows of it, found that counsel information stops in 2020 because Harvard's metadata stops in 2020, and cancelled the product. We are thanking them for the infrastructure, not for a product, because there is no product. The measurement cost a day and would have cost far more later, and the people wrongly listed would have paid part of that bill.

The Wikimedia Foundation's Commons carries the 71 portrait attributions on our /humans gallery. Every face there is public domain or openly licensed.

Three records requests are drafted and pending with state bodies. None has been granted, and we are not going to describe a pending ask as access. When a source looks closed we check whether the restriction is about the requester's purpose before we record it as blocked, and we go through the front door rather than around it, every time.

And the group we owe the most to did not volunteer for any of this. Hundreds of thousands of licensed professionals are in our directory because a public register lists them, not because they asked us to. What we owe them is not a thank-you. It is a removal path that reaches a human, costs nothing, requires no account, and survives the next rebuild. That last clause is the one we got wrong and had to fix. It is at /directory/claim and it works today.

The code we did not write

The visible surface of this product is largely other people's work, given away. Next.js and React and Tailwind. Radix UI for the primitives that have to be accessible and usually are not. three.js with react-three-fiber and drei, and GSAP, for everything that moves. MediaPipe for the on-device model runtime. react-markdown, remark-gfm, lucide-react, d3-hierarchy, simple-icons. TypeScript, Vitest, Playwright, ESLint and jsdom, which is to say the entire reason the previous section could be written with numbers in it. OpenTelemetry for the tracing we have wired and not yet turned on.

There is no LICENSE file in this repository. We take from that ecosystem every day and we have not done the paperwork that would let anybody take anything back on the same terms. Writing that down is not a fix. It is a note we can no longer pretend we did not write.

The people we learn from at a distance

Our engineering language comes from public teaching we had no part in. "Code like bacteria," the preference for small self-contained modules anybody could lift without importing our whole world, is Andrej Karpathy's idea, learned from what he publishes. Our standard for candour is Charlie Munger's, and our Rude FAQ is written in his spirit. Jeff Dean has been a long-time mentor to our founder. Luiz André Barroso, who died in 2023, was his mentor and his last boss at Google, and is the one genuinely personal relationship in this paragraph. Jim Simons is the north star for how the parent fund thinks.

None of the people named in this section hold any role at 🤫 hussh, HushOne, Inc., or Hushh Technologies Corporation. Naming them shares our admiration and what their public work has given the world. It implies no affiliation with, endorsement by, or sponsorship by them. The same is true of every agency, regulator, and open-source project named above, and of every organisation that employs one of our advisors: those advisory relationships are with individuals, and their employers have nothing to do with us.

The ones who are not in the log

Eleven advisors and one board member are listed publicly on our team page. We are not going to attach an invented anecdote to each name to make this section longer. They read things nobody paid them to read, and the detail is theirs to tell rather than ours to embroider.

And one more fact, stated because it changes how you should read everything above. This project is funded one hundred percent today by a single American citizen, our founder, as research and development given to America for free. Not a startup that also does good. That is why there is no pricing page aimed at a citizen and why there never will be one. It is also why this post exists: he asked for a gratitude post, and he asked that it not be a victory lap.

What would make any of this true

Gratitude is cheap to write. So here is the test we would rather be judged on.

Send us a removal request next month. If your record is gone, and still gone the month after that, when the registry we built it from still lists you and the rebuild has run twelve times in between, then the thanks in this post were real, because the only thing holding your record out is a promise somebody turned into a test.

The full record of what shipped is at /releases, and the people we celebrate are at /gratitude. Both are more reliable than this post, because both are generated from things that happened.