Luis Alfonso Reyes

Cross-border law is scattered across government sites, languages, and print-first documents that change every week. There is no single place to search it.

General AI will still produce a fluent answer when its source is missing. In law, an invented citation is not a minor defect; it can become malpractice.

I collected and cleaned 736,833 articles from 6,537 laws across nine countries and international bodies, plus 321,589 court decisions. Semantic indexing lets a question in Spanish retrieve the Czech ruling that answers it.

On top of that sits an assistant that drafts documents. Behind it sits the part that matters: before any answer leaves the system, every citation in it is re-checked against the documents that were actually used to write it. A citation that cannot be confirmed is flagged, not shipped.

I built all of it — the database, the back end, the interface, the billing, and the jobs that run every day to keep the law current.

A live business with paying subscribers since October 2025: 1.47 million records kept current and six tools in one product. Its data feed has served 1,048 requests from 15 external clients. One engineer, start to finish.

The company builds and restores historic windows. Its whole operation lived in thousands of loose spreadsheets and CAD drawings across shared folders. Nobody could answer “did we make money on that order?” without opening six files, and by then the material prices had moved. The data is commercially sensitive, so putting it in someone else’s cloud was not an option.

One application covering sales, production, purchasing, warehouse and shipping, with a catalogue that knows the materials and the hours each window needs. It reads their existing spreadsheets and drawings, so nothing had to be retyped.

It installs on their own server with a single command and upgrades itself without losing data — because I am not there to run it for them.

The number that mattered to them was cost, so I made cost behave. An order’s cost updates while the order is open and locks the moment materials are bought. A price change next quarter can no longer rewrite last quarter’s profit — which means that, for the first time, their margin history is worth arguing about. Running on their own hardware today.

The instructor had recorded a course and wanted to sell it herself. Every ready-made platform takes a large cut, and none of them stopped a paying customer from downloading the videos and passing them around. She also needed the whole thing in two languages, and needed to change her own class schedule without calling me every week.

A 31-page site in Czech and English, plus a members’ area. Customers sign in with a link sent to their email — no password to forget, lose or share — and each link works exactly once.

The videos are never handed over directly: every play request is signed and expires, so a copied address is dead on arrival. I tested that the way it should be tested, by trying to steal a video myself and confirming the server refused.

Payments run through the Czech bank gateway, and the customer’s language survives the trip to the bank and back because it travels inside the part of the request the bank itself signs — so nobody can tamper with it on the way.

24 lessons behind a lock that holds. She runs her own schedule and her own campaigns from a spreadsheet and the site follows, which is the difference between a website and a dependency. The bank’s compliance review came back with two conditions before go-live; I closed both, which meant writing her terms of sale, refund policy and complaints procedure. That part is not code, and it is why a lawyer who builds software is useful.

She had a printed brochure and no presence online at all, in a practice where clients start by searching. She also had no way to answer, at two in the morning, the only question every visitor actually arrives with: what does this cost.

I built the site out of her own brochure — its colours, its typefaces, its prices — so a client who has both in front of them sees one firm and not two. It carries an assistant that answers the common questions about registering a trademark, and an intake split into two short steps instead of one long form.

It ships in two complete looks, light and dark, chosen by the visitor: every colour is defined by its job (“this is a background”, “this is a heading”) rather than by its value, so switching is a swap and not a second website to maintain.

The first draft repeated her six practice areas three times over. I cut the page by roughly a third — 1,027 words down to 809 — while adding the price list that had never existed. Shorter, and it finally answers the question people came with.

Search by meaning has a blind spot that nobody advertises. It finds documents that sound like your question. It cannot find the landmark ruling that every relevant case cites but that shares almost no words with what you typed — and in law that is very often the one you needed. I wanted to know whether the citations themselves could close that gap.

I rebuilt part of my corpus as a network of who cites whom: 622 court decisions, 4,063 citations between them, and 4,292 pieces of legislation with 1,393 internal cross-references. A search now runs twice — first by meaning, then by walking outward along the citations from whatever the first pass found.

On a test question about enforced disappearance, ordinary search returned five cases. Walking the citations surfaced two landmark judgments it had missed completely. The same exercise exposed 24 corrupted citations that had been sitting unnoticed in the source data — a data-quality problem I would never have seen otherwise.

The offer only works if it is instant, and she was doing it by hand: reading every comment, opening each person’s messages, pasting the link. On a post that lands, that is a hundred conversations in one evening, and the people she got to the next morning had already moved on.

She teaches most of the day, so the hours when a post is hottest are exactly the hours she cannot answer. And she did not want to call a developer every time she changed a word.

Everything she controls lives in her own spreadsheet: the trigger words, the private message, the public replies — the bot rotates between them at random, so the comment section does not read like a machine — and an on/off switch per campaign. It re-reads that sheet every five minutes. Nothing to restart, nobody to call.

The harder half was deciding which posts to watch. Instead of registering each campaign by hand, the bot reads her captions and recognises her own call to action: the word slovo followed closely by a trigger word, which is how she has always written it. Tested against 47 of her real posts, it found both campaign posts and left alone the other 34 that merely mention the same words — including the trap, a post that says “write to me in the comments” and also says the word, but never asks anyone to comment it.

It is built to fail closed. With no trigger words configured it does nothing. If it cannot tell which posts are campaigns it acts on none — it never falls back to answering everything. Words match at word boundaries, so it catches the Czech inflections (dechu, dechem, běhám) without firing on průběh, where somebody was simply asking a question. It never replies to itself, and it never goes back to message people who commented before the campaign existed.

The message text is never written to the log. Her clients write about their bodies, their health and their appointments; that is sensitive data under GDPR and it has no business sitting in a server log. Comment text is kept, because it is public and it is what you need to debug a trigger word.

Live since 30 July 2026. In its first six weeks it answered 161 people with 173 private messages and 174 public replies — all within seconds, none typed by her. It discovers a new campaign within fifteen minutes, so she can launch without calling me. The goal was not only to remove her repetitive work; it was to remove me as a dependency.