711 239 085 info@gilsys.com
es ca en
Abstract background

Near Vic (Osona), in the province of Barcelona, Spain

Custom development and software with artificial intelligence

We automate processes and build applications that interpret information and make decisions, integrating commercial AI models (ChatGPT, Gemini, Claude) or open-source models on private servers.

We offer advice on how to integrate AI

In an AI consulting engagement we analyze the company's processes and data and propose where to start, with an assessment and a roadmap.

See AI consulting

Integrating AI into a company is the responsibility of a software development company

Would you put your family in a car built with AI by someone who knows nothing about mechanics?

Developing a basic tool with artificial intelligence seems simple; making it secure, maintainable, and integrated into the company is not. The difficulty begins after the first prototype: getting that AI to read the ERP data, write to the program the team uses, and respect each person's permissions.

That work is no longer about configuring a tool, but about programming. That is why an AI development company can carry out both the assessment and the implementation: the same team that studies where AI fits in develops the software that integrates it, with no intermediaries and no information lost between those who analyze and those who develop. Afterwards, that same team maintains and further develops it, because it knows at all times how it is built: it is not a black box that simply works.

More and more companies present us with a prototype built with AI that they consider almost finished. On review, what usually happens is that what can be seen works and what cannot be seen is missing: the data structure, the permissions, and error handling. In those cases we review it, decide what can be reused and what needs to be rebuilt, and plan it as a first phase with its own scope and budget.

"AI gives programmers superpowers, but judgment cannot be automated."

Jordi Gil, founder of Gilsys, in an interview with MetaData

AI tools we have developed

Complete tools in which AI is the core component. AI integrations inside an existing program or process are explained in each of the sections above.

Private AI on the documentation of a social organization

A non-profit organization uses AI to query its case files, internal protocols, and codes of conduct. The open-source model runs on the organization's own server alongside its document management system, ownCloud Infinite Scale, and each team can only ask about its own workspaces. Nothing leaves the organization. See the case.

Public tender finder with agents and Jev

An agent periodically reviews newly published public tenders, Jev decides whether they fit the company's profile and to what percentage, an automation downloads the tender documents, and the person in charge of tenders decides whether it is worth bidding. See how it works.

In-house SEO tool

Every day it gathers data from Search Console, Google Analytics, and PageSpeed, diagnoses each page of the website, and compares its content with that of the companies ranking ahead in Google. The AI accesses that data through the tool's own MCP server, with permission levels (read, propose, apply), prepares change proposals, and a person decides which ones are applied.

Code review agent

Once our team has validated a development, an in-house agent, configured with our guidelines, carries out the final review of the code and its security, and its report is checked by hand before publishing. See the security review.

How does AI connect to the company's software?

AI adds value when it works with the company's real data, and that requires programming the connection to each system.

Inside the programs already in use

AI is integrated into the ERP, the CRM, the document management system, or the custom application, and it looks up and records information in them. Nobody has to copy and paste data into a separate window.

With each user's permissions

Each person gets answers from the information they already have access to, not from the whole system, and the AI only carries out the actions it is authorized to perform.

Also with legacy programs

There is no need to replace a program that is years old to take advantage of AI. On top of databases such as MySQL, Microsoft Access, or an AS400, we program an intermediate layer that reads and writes through whichever channel each one supports: an ODBC connection, an API, or the database itself. This is systems integration work.

From ChatGPT or Claude, with an MCP server

MCP (Model Context Protocol) is the standard these assistants use to connect to other programs. We program the server so that it offers only the queries and actions that have been agreed, and the team works with ERP or CRM data from the assistant it already uses. With a commercial assistant, the information queried reaches its provider; if it cannot leave the company, the same server connects to a model installed on private servers.

Which AI model is used in each project?

It is the first decision and it shapes everything else: the cost, the speed, and what information leaves the company, but it does not have to be final.

A commercial model

ChatGPT, Gemini, or Claude. They are usually the most capable, are paid for by usage, and require no maintenance. The trade-off is that the information sent to them leaves the company and the price depends on a third party. It is the right option when the information being processed is not sensitive.

An open-source model on private servers

It is installed on a server the company controls, and the tool belongs to the company. Neither the documents nor the questions leave the organization, and the cost is that of the server, not of usage. The trade-off is that it has to be sized and maintained: this is what we implement in a private AI project.

A gateway for switching models

We connect the application to a model gateway, for example LiteLLM, instead of to a specific provider. This makes it possible to switch language or reasoning models, split traffic among several, or replace one that has gone up in price without modifying the code or redeploying. When a new version of the same model is released, for example of Claude Sonnet, it is tested with real cases and adopted while keeping the same instructions, so the tool improves without changing its behavior. In a field where a better model appears every few months, this is an important advantage.

The same for image generation

To generate or edit images we use platforms such as Fal.ai, which give access to many image models through a single integration. Several are tested with real examples, and we switch to the one that gives the best result or the best price without rebuilding the tool.

Information should be as well organized as possible before integrating AI

AI inherits the company's data

If information is duplicated or incomplete, an AI system returns duplicated or incomplete answers, and presents them as true. For example, if the same customer has been registered three times with the name spelled differently, the AI treats those records as three different customers. With well-organized information, an AI project moves faster and costs less.

  • The same models detect duplicates that are not identical and values that do not add up, and a person decides.
  • A system with rules for each type of information: a product information management system, such as Pimcore, for the catalog or a document management system for files.
  • Before the AI project or in parallel.
See how information is organized and validated

How we work on an AI project

From Sant Quirze de Besora (Osona, Barcelona), we work for companies throughout Spain: in-person consulting only in Catalonia, and remote consulting, development and automation for the whole country.

First meeting, free of charge

We learn about the company's situation: what problem needs solving and which programs and data are used today. We also assess whether AI is the answer or whether traditional programming is enough.

Consulting or straight to the process

If it is not yet clear where to start, AI consulting analyzes each area and prepares an assessment and a roadmap. If the process to be improved is already clear, we move straight on to defining it.

Scope and budget of the first phase

We define the requirements and fix the scope and budget of a first phase; the following ones are planned based on its results (phased development). The budget also determines, based on the conversations held, whether the process requires human supervision.

Development in three environments

We set up a development environment, where we test; an integration environment, where the company tests; and the production environment, for the real process. A basic feature can be developed in a week, but conceptualizing it, defining the requirements, testing it, and putting it into production usually takes longer, depending on the task being automated.

Go-live and evolution

With the process up and running, we compare the result with the starting point and plan the next phase. During the maintenance phase we can carry out periodic updates of the tool, as well as switch to new versions of the AI models.

Custom software development powered by AI

How does Gilsys use generative AI for programming?

The architecture, security, and business model are designed by our engineers. On that foundation, generative AI supports us in several phases of development so we can deliver more robust, better-tested software in less time.

  • Working prototypes from the start of the project
  • Automated tests before each version is published
  • Final code review with an in-house agent
See how we program with AI

Frequently asked questions about AI in business

What can be automated with artificial intelligence in a company?

Many repetitive tasks involving text or data: summarizing contracts and reports, classifying documents or products, serving customers, or managing appointments with a conversational assistant. AI has advanced a great deal and the list does not end here. With more than 20 years of programming experience, we study any idea the company puts forward and assess whether it can be automated.

Are there processes where you do not recommend using artificial intelligence?

Yes: problems that require an exact result or strict compliance with a set of rules. For example, seating event attendees at tables with dozens of incompatibilities and preferences is a combinatorial problem: an optimization algorithm based on graph theory finds an arrangement that meets every constraint or proves that none exists, whereas a language model, being probabilistic, may break one of them without noticing. AI is well suited to interpreting ambiguous information and reasoning with it; when it comes to calculation or formal logic, we apply traditional programming, and both can be combined in the same project.

What is the best AI for a company?

There is no single one. It depends on the task, on whether the data can leave the company, and on how much each use costs: a commercial model such as ChatGPT, Gemini, or Claude is usually the most capable for writing, summarizing, or reasoning; an open-source model on private servers, for sensitive information; and a model that only decides, such as Jev, for classifying many cases quickly and cheaply. What is more, today's best model will not be the best in a few months. That is why we recommend connecting each application to a model gateway, which makes it possible to switch models without rewriting the code. That is how we work at Gilsys: what has been developed is reused when a better model appears, instead of being rebuilt.

Do you build chatbots for businesses?

Yes. A chatbot that answers with the company's data is built with the same components as a private AI built on documentation: a vector database with the documentation, a language model, and an interface. It connects to the CRM, the ERP, or the online store when it needs to look up live data, and it can be used by the internal team, inside an application, or by customers.

What does the price of a custom AI project depend on?

Mainly on three factors: which model is used (a commercial service such as ChatGPT, Gemini, or Claude, or an open-source model installed on private servers), how many company systems it has to connect to (CRM, ERP, online store, custom applications), and how much in-house data it has to work with (documentation, ERP data). The model accesses that data at the moment it answers: it almost never needs to be trained. That is why each project is quoted separately; the quote is free and with no commitment. The remaining factors are explained in what determines the price of custom software.

Who owns the development and who pays for the AI usage?

The code belongs to the company: it is delivered documented, in its repository, and deployed on its servers or in its cloud. Use of the AI models (OpenAI, Anthropic, Google…) is normally contracted in the company's name, so the invoice comes directly from the provider, with no intermediaries. If the company has its own technical team, internal APIs can be left in place so that it can extend the application without modifying its core.
Abstract background

What task is repeated every day in the company?

In a first conversation, free of charge, we analyze what it is and what data it relies on today, and assess whether AI can solve it.