What we make

Every organization has one thing only it can do. Everything else can be built.

Every organization needs the same unglamorous machinery, and none of it is the reason anyone got into this. We build it, run it until the owner's people can, and stay out of the way, so an owner's hours are spent on the part nobody else can do.

We build the thing that is needed, to the owner's requirements. The owner holds it outright and can look at every part, and we will run it as well, for as long as the owner wants. What is below is three situations where it mattered, and then what has already been built. All of it is evidence of what we can make, not a menu to choose from.

These are past situations, not a list of services. No two situations are alike, and the next one will look like none of these. What carries over is how they were approached.

A redemption nobody had planned for. A fund faced a major redemption. It had one of the best administrators on the street, and still nobody trusted the figures, least of all for side pockets that carried future funding commitments. The systems had been built to collect fees. Nobody had imagined paying investors out. Reconciliation reports had stopped arriving, and a month's NAV took two months. It was settled when people ran out of objections, not when it was right.

A redemption is a well understood process, and it was never the difficulty. The difficulty was the partners' fear, and it was reasonable. So the work was to develop a process they could believe, win their agreement to it, and then carry it out without a single error. The new tools had to be built and delivered while the day-to-day carried on.

The administrator had the records to defend its work, and its position was a fair one: "I know what I was told and what my systems captured, and I will work very hard on the special case, but the risk is not mine." It is the client's balance sheet on the table, and a wrong number can be larger than the manager's own net worth.

Most relationships with a third party's bench are like that. A service defines its role and limits it: "I will come to the office and empty the bins, and keeping the office clean is not my responsibility." That is a fair bargain, and it leaves the risk where it was. The auditor is no answer either. Audit standards and the needs of a running business are two different things, and an auditor is far too expensive to run operations.

Nobody but the business signs off for the business. However large the bench it hires, the risk stays its own. So the work of being right has to be done inside the entity, for the whole outcome and not for a defined slice of it. That is where we work.

The founder of Ploutos Services was that fund's chief operating officer. More than five years of history were rebuilt from the records. The errors were found, every allocation made by hand was traced and understood, the missing data was filled in, and money was returned to investors that the corrected books showed was theirs. The offering memorandum was redrafted and the redemption notices redone. Paper wire orders gave way to payments sent over SWIFT.

The documents required the money back with investors in three weeks, which left one week for the NAV of a multi-strategy fund holding a wide variety of investment products. It was done in a week, fully reconciled, and signed off by everyone. Two months had become one week.

Then it was handed over. The people who took it on did not need to understand everything its builder understood. The process carried its own checks, more extensive than any they had run before, and the checks made up the difference.

Thousands of derivatives, and nobody who could check the bank. Twenty years ago OTC derivatives became fashionable at a large fund, and soon it held thousands of them. Some settled monthly. Others exchanged margin with the bank every day. The fund could not do its monthly resets, and when the bank called for margin it had no view of its own to set against the bank's. Two answers were on the table. One was to leave twice the required margin at the bank, so that the bank would never need to call: asked for ten, post twenty. The other was to hire a treasury specialist at five million a year.

The team that reached those two answers was competent, intelligent and had worked very hard. Its case was that it had tried, and that its conclusion deserved respect. Years of managing people teach that most arguments rest on what someone said and what someone heard. The meetings generate their own issues, and little of it survives sunlight. That does no harm while the machinery is well built. It offers no cure when the machinery is broken, because a broken machine cannot fix itself. The assumption nobody had examined was that the bank's figure was the figure.

So the first step is to step outside experience. "I sat in the car, I pressed the pedal, and the car did not move." Experience says a car moves when the pedal is pressed, and it is nearly always right. It is no help on ice. Plenty of the cars in the ditch are four-wheel drive.

Neither answer was taken. Margin is not a fact to be computed. It is one party's view against another's. So the fund formed its own view every day. The trading desk marked every position with a justification that could be taken to the bank, and the fund made its case and stood its ground when the numbers differed. The money moved over the SWIFT process already built. Two recent graduates ran treasury, and had treasury management on their résumés. The monthly equity resets had been two months of argument that ended with the same errors it began with. They were now checked and reconciled daily. From the bank's high-risk category, the fund became one of its best-managed clients.

A change the old system might not have survived. A school ran on a roster system of its own, begun fifteen years ago in PHP 4 and tied into Google. It needed a great deal of hand-holding. Its teachers and families came and went, and they used their own email addresses. Giving each of them a school account is close to unmanageable for an audience that fluid: before long there are two thousand accounts and forty in use. So everything was built around personal addresses.

The replacement was already being built, as a successor to the old system and not as a rescue. It was a completely new system, built in four months by one person working with AI models. It took the ideas and fifteen years of data from the old one, and not one line of its code. During that work an AI model found, by trying it, that Google had changed its rules. Its classroom would no longer accept a teacher on a personal address, and it would no longer send a class invitation to a student's personal email. Much of what the school had relied on for years had quietly stopped being allowed. The same model laid out the alternatives, and the one chosen was built. One part of it is told in detail.

The school opens its year on the new system, stronger than before. Most tasks are self-service, and it is documented far better than what it replaced. What would have happened had the change been met in the old system, a week before opening, nobody knows. Survival in the wild takes a good deal of luck, and not everyone is lucky. What method does is put a venture where the luck can be used: the new system was already being built when it turned out to be needed. When a road is blocked, the question "why does it not work?" has no useful answer. One finds another road.

The point is not which problems were solved yesterday. It is the thinking that solved them. None of these were new problems, and none was complicated. Every one had been solved somewhere before. The failure is in the method: drawing water from a well by jumping into it, when a pulley has been hanging there all along. What is rare is looking at a business from first principles, without bending its problem to fit the answer one already knows. None of it rested on knowing the domain beforehand: each was new ground when the work began. The opposite of the method has a name in Hindi, jugaad: starting the car by banging on it. It works, which is the funny thing, and it is found everywhere. People bend the problem, and AI models do it too. So the understanding is worked out once and built into the machinery, with checks that catch whoever bends it next. It is the difference between "I thought" and "I checked".

The first two took a team of programmers. The third was one person with AI models, which is what the rest of these pages are about.

The things below look unrelated: a set of books, a school, a family tree, a building. They are the same problem. Until recently no one person could have built them all, because each needed knowledge that took years to acquire. With AI models, knowledge has stopped being the boundary. That cuts both ways: anyone can now do our job, and we can do anyone's. What stays scarce is what always was, which is seeing what the problem actually is. Whoever was ahead because of what they knew now stands on new ground. Whoever was ahead because they could see the problem is still ahead. Reconciling a bank account, tracing a family tree and running a building turn out to be one skill.

  • Books that keep themselves Statements arrive as files and become records. The ledger is generated from those records, in any currency, across any number of entities, and it can be rebuilt from scratch to prove it.
  • Several worlds that never mix The business, the family office, a trust, the household, on one set of machinery, each seeing only its own. The separation is enforced by the database itself, and when it fails it fails to an empty answer, never to somebody else's rows.
  • Documents that file themselves Forward an email and it lands where it belongs: on the calendar, in the books, in the record. The inbox stops being a to-do list.
  • One file, referenced everywhere Store a document once and point at it from twenty places. Delete it in one, and the other nineteen still work.
  • A private library of an organization's own records Agreements, licences, registries, properties, as pages people sign into with the accounts they already have, each seeing only what is theirs.
  • A question box over an organization's own material Ask in plain words and get an answer that cites the page it came from. Each reader is answered only from what that reader may see.
  • A public site that answers its own questions Visitors ask, and it answers from the site's own facts, not from the internet's. The box near the foot of this page is one. It is tested against a scored question set before a change goes live.
  • A school that runs itself Families enrol, pay and see their own children's classes, calendar and reports. Staff see the school. Nobody retypes anything.
  • A building that runs itself Lights, shades, climate and the alarm, on one door, with no vendor's app in the way. Arrive after dark and the lights are already on.
  • A building's own readings, kept by its owner What the building produced and what it drew. The past is copied from the manufacturer's cloud, and from the day we arrive it is measured on site, so it can be put next to a bill and argued with.
  • A family tree that answers questions Not a chart to squint at. Ask how two people are related and get the answer in a sentence.
  • A family's photographs, on its own storage Catalogued, searchable, and not renting space from anyone. The bill stops being a subscription and starts being a disk.

The ground keeps changing, and the old answer is usually still on offer.

Documents are read by machine. Statements, invoices and spreadsheets arrive as files and are read and modelled as data. The old answer was to hire people to read them.

A page that expires. The old answer was a PDF emailed with a password, which then lives in an inbox for ever. Today the reader gets a link to an encrypted page that is hard to copy and stops working after a set time.

Records are kept on the owner's own disks. The old answer was to hire a records warehouse. Today they sit on disks the owner holds, on more than one continent.

Layers, because everyone works somewhere else now. The exposure to the internet is far greater than when staff sat in one office. So there is a private network, protection of the data itself, and more than one check on the way in. The library can keep a file sealed, not merely behind a sign-in.

Our own certificate authority. The old answer was a password shared between systems, which lets any one of them pretend to be another. Here every application is issued its own certificate, renewed without anyone touching it. Few organizations of any size attempt that. It is ordinary here.

Every layer can be replaced. In 2012 nobody asked about passkeys. Today they have to be supported. Whatever is built has to let the next such thing in without being rebuilt around it.

One of the things made is running where anyone can look at it. vidyalaya.org is a school built this way. Families enrol and pay online and read their own children's classes and reports, and the question box on the front answers from the school's own facts without anyone signing in. Ask it something the school would know, and something it would not.

What can be built this way is growing. An AI model, used well, can take over the repetitive work that today needs a person in a chair, so each year more ideas become practical for one owner to carry out without a department. Asked to find the right place in a PDF and put the founder's signature there, a model simply did it. Two years ago that was an errand for a person. It changes what is worth buying, too. In 2012 the right answer to a large accounting problem was to pay for a large accounting suite, licensed by the seat, and bend the business to fit it. Today the right model, told exactly what the books must do, can write the ledger this owner needs.

A model can be told what to do, and it listens. It has nothing invested in the old way: no pride, no career, no years of work that a new way would make worthless. That is a great deal to ask a person to set down. It still mishears, which is what the law and the checks are for.

Who we are, who this suits and the three ways to work with us are on the front page. For a reader who wants the machinery, it is written down in detail.

Start with a conversation