Our work: what we built

Mesterséges Intelligencia Központ Kft. is a software development company based in Budapest, Hungary. What we built: an ERP with invoicing and CRM, a program that pseudonymises legal documents, a database sync service, a school laptop fleet manager and a 3D game with a live AI commentator.

Every system described here exists and works; where we are still putting one into operation, we say so. We do not disclose the names of our clients or of the products. If you need a system that is not listed here, the end of this page also lists what else we build.

An ERP with invoicing and CRM

Challenge. The daily work of a Hungarian small or medium-sized company is typically spread across separate tools: spreadsheets, e-mail, a separate invoicing program and the bank's interface. The goal was one system built around the company's own processes, which several organisations can use without seeing each other's data.

Solution. We are developing a modular, multi-tenant enterprise resource planning (ERP) system:

• calendar, tasks and approvals, and background jobs;
• mail merge and a built-in mailbox for the existing e-mail accounts;
• sales and CRM, procurement, education administration, and HR with AI-assisted analysis of CVs;
• finance: outgoing and incoming invoices, bank, transfer batches, payroll and general ledger;
• reports and charts from every table, and a docked AI assistant that can also be controlled by voice.

Invoicing issues invoices reported to the Hungarian tax authority (NAV) through the invoicing program, with a PDF preview and cancellation. Bank reconciliation matches each payment to its invoice by the transfer note, the invoice number or the payer's account number, and posts the payment back; incoming invoices received from NAV go into transfer batches after approval.

In the CRM module, partners can be organised into chains, contacts have decision roles, sales pipelines move on kanban boards, stage changes can be traced, tasks come with reminders, artificial intelligence summarises the partner's timeline, and e-mail and calendar are synchronised to the partner record.

One login gives access to every module: the central sign-in works with a password or a Google, Microsoft, Apple or GitHub account, with two-factor login, organisations, invitations and permissions.

Result. The system is live at two organisations; invoicing is live, and the CRM module is being rolled out. We put this experience to work in business software development.

A document anonymizer for legal documents

Challenge. A contract, a court filing or a judgment often needs to be shared or processed with artificial intelligence, but no personal data may leave. Redacting by hand is slow and error-prone: a name also appears inflected, as initials or as a signature.

Solution. A Windows application that runs entirely on the user's own computer and pseudonymises Hungarian legal documents (PDF and Word):

• it finds every form of the parties' names: inflected forms, maiden and married names, initials, signatures in capitals and names inside e-mail addresses;
• a Hungarian name-recognition model running locally finds the other names, organisations and addresses, and in our measurement it was 10–13 points more accurate than generative models;
• it offers four kinds of replacement, including court-role labels and numbered labels, and shows a review sheet before saving;
• it rewrites the file while keeping its format, including headers, footnotes, comments and metadata.

Result. A second software library checks the finished file again, and if even one original name remains, the program does not save. The application can be installed, and the data never leaves the computer.

A 3D browser game with a live AI commentator

Challenge. A company website usually says that the company understands artificial intelligence. We wanted to show it. The game had to work both at a trade show stand and on the website, for business visitors who are not gamers: with nothing to learn, no registration and no start button. And during play the AI had to genuinely react to the player instead of drawing pre-written lines at random.

Solution. We built an endless runner in which the player is the T-rex itself. WebGL draws the scene in real time with three.js, in the browser; the controls are a single button. The game sends its important moments (a lost life, a record, a level-up, a new landscape, a cleared chasm) in a summary report to our AI assistant, which is built on a large language model and speaks to the player live on that basis. The assistant only talks: it never interferes with the state of the game.

Result.

• For the big moments, the commentary appears in roughly 1–2 seconds.
• Our AI assistant speaks 14 languages, always the player's language, and uses the formal form of address in every language that has one.
• It always answers a question the player types, and at the end of the run it sums up the round with at least two numbers.
• The same game runs on desktop and on phones.
• The game is live: anyone can try it on our homepage.

We bring the same know-how to our game development and our AI chatbot and agent development work.

An AI-based production line for game graphics

Challenge. A game needs many graphic elements in a consistent style: trees, shrubs, rocks, landscapes, obstacles. Each of them has to be made at true scale, with a transparent background and an exact outline, because the game's collision detection relies on it too. By hand this is slow, and simple image generation does not give a consistent style and size.

Solution. We built a six-step production line into our admin interface:

1. Order: a short description of the theme.
2. Plan: a language model writes the plan of the world, covering mood, colour palette, weather, the kinds of objects and the drawing instructions.
3. Style sheets: selected sample images fix the style and the scale; the image model receives them as references.
4. Drawing: a generative image model draws the objects, one sheet per kind, with several sheets in parallel; the machine cuts each sheet into individual objects, checks transparency and the complete outline, scales to the planned height, fits the colours to the palette and creates a normal map for the game's lighting.
5. Selection: a person decides on acceptance.
6. Release: only released images go into the game.

A scheduler produces a saved order on its own within a set time window and a preset AI call budget; selection and release stay manual even then.

Result. The game's landscapes and obstacles were made on this production line. An obstacle's drawing, icon and collision shape come from the same production data: the collision follows the pixel-accurate outline of the image. Nothing gets into the game without human approval.

Function point sizing with artificial intelligence

Challenge. A software offer is only as good as the estimate behind it. An estimate in hours depends on the estimator and the team, so two offers are hard to compare, and the client cannot see what exactly the developer is pricing.

Solution. We measure the functional size of a development in COSMIC function points (ISO/IEC 19761). The method counts data movements (Entry, Exit, Read and Write), each worth one function point, so the size is independent of the technology and of who does the development. For the measurement we built our own AI-assisted workflow:

• we break the development down into groups of functions and describe each with the same template: goal, actors, data, functions, AI functions and what is out of scope;
• several language models count every group of functions several times; the runs build a shared dictionary of the names of actors and data, and their spread shows the uncertainty;
• a reconciliation round ensures that every functional process counts exactly once in the project, even if it appears in several descriptions;
• the existing system is described separately, so only the new part counts towards the size;
• the measurement also flags what the functional size does not express: a custom interface, graphics, real-time behaviour, a complex algorithm, data migration, non-functional requirements;
• the automatic merges can be overruled by hand, and the finished measurement can be locked as an unchangeable version, which becomes the basis of the contract.

Result. The offer and the schedule are based on the measured size, and two offers can be compared on the same scale. The size is not a gut-feeling estimate but an itemised, traceable list: which process, with which data movements, worth how many function points. Our Pricing page explains the method.

A sync service from MS SQL to cloud PostgreSQL

Challenge. The data of an MS SQL database running on a company's own server had to reach a cloud system (for a new application and for reporting) without loading the source system and without writing anything into it.

Solution. A program running as a Windows service that works with read-only rights:

• for each table it picks the cheapest way to detect changes, and it finds deletions by comparing keys;
• it reacts to a change within 1–3 seconds, creates the structure of the target database itself and follows its changes;
• sensitive columns can be excluded, bandwidth can be capped, it reconciles at night and reports row counts;
• it is installed with a single installer and stores its access credentials encrypted.

Result. Its operation is checked by automated tests and by a real MS SQL and PostgreSQL test environment. When several systems need to be connected, that is process automation work too.

A laptop fleet manager for schools

Challenge. About a hundred Windows school laptops at several sites, with no domain and no device management system: installing software, updating and keeping accounts in order was done machine by machine, by hand.

Solution. We built our own fleet management system:

• the agent on the laptops only connects outwards, the central server is managed from a web interface and keeps an inventory of every machine;
• commands reach the machines in stages, with a kill switch and a log;
• it installs and removes software from a catalogue and reports which machine has drifted from the required state;
• it manages uniform local accounts and stored administrator passwords, and updates itself in stages, with automatic rollback on failure.

Result. The fleet runs on it. The decision to move to Windows 11 was also made on measurements: on the results of benchmarks run on the laptops.

AI testers for web applications

Challenge. A web application needs retesting after every change, and manual testing is slow and repetitive.

Solution. AI agents go through the given web applications in a real browser, complemented by automated and accessibility checks. A person approves each finding, and only then does it become a ticket; the agents learn from the approval decisions. Runs can be scheduled and only revisit what has changed; pages are also checked in a phone view.

Result. It runs as an internal tool, for testing our own developments.

A hiring test for people who work with AI

For our own hiring we built a half-hour online test that measures how a candidate works with artificial intelligence. It has four timed parts: bug hunting in a deliberately faulty demo application, a conversation with a client played by artificial intelligence, reasoning tasks and an experiment with a system whose workings are unknown. The AI pre-assesses the answers, a person makes the decision, and candidates' data is handled under the GDPR, with a retention period, export and deletion.

Smaller products of our own

• Dictation program: types speech into any application on Windows and macOS; speech recognition runs locally, on the computer's graphics card.
• Poster campaign verification: a staff member's photo and location confirm each poster put up, with an overview on a map.
• Trading card manager: the phone camera recognises a card by image matching and text recognition, and the collection and the decks are managed in one place.
• Our previous website: an AI agent restyled the page live for every visitor, while another one talked with them, voice input included.
• This website for search engines and AI assistants: in 14 languages, with a Markdown version of every page, machine-readable company data and a sitemap that marks the language versions.

What else we build

The following are not finished projects but systems we would be glad to build:

• extracting data from documents such as invoices, contracts and delivery notes with artificial intelligence: AI integration;
• a customer service chatbot or AI agent that answers from the company's own documents: AI chatbots and agents;
• a webshop with custom pricing, connected to the ERP: Webshop development;
• an iOS and Android app with a back end and an admin interface: Mobile apps;
• a branded game or interactive experience for a trade show or a website: Game development.