The context
The Instituto de Gestão Previdenciária e Proteção Social do Estado do Pará, the state's social security institute, had already been pursuing a digital transformation journey across its systems and services.
Within the Procuradoria Jurídica (legal counsel's office), however, the Coordenadoria de Execução (enforcement unit) faced a challenge of a different kind: improperly paid amounts had to be identified, analyzed and collected before the statute of limitations ruled out their recovery.
Unreported deaths, changes in beneficiaries' status and other events could result in payments that later had to be recovered.
The process was predominantly manual and reactive. There was no structured view of the collection portfolio, viable and unviable cases landed in the same queue, and a significant share of the legal team's energy was consumed by repetitive tasks before the actual legal work could begin. The longer each case took to analyze, the greater the risk that recovery would become time-barred.
Before AI, the process had to be understood
The first decision was not choosing a language model.
It was understanding how debt recovery actually happened.
Vibe and IGEPPS applied Lean Thinking as an improvement philosophy and used Value Stream Mapping (VSM) as the tool to see debt recovery as an end-to-end value stream.
The goal was to make visible what had previously been scattered across departments, documents, systems and individual decisions.
The mapping helped identify:
- where time was being lost;
- where information disappeared;
- where there was rework;
- where decisions got stuck;
- which steps depended on incomplete data;
- which activities truly required legal analysis;
- and where technology could actually create value.
The work also revealed that some of the main bottlenecks came before the legal analysis: locating the debtor, data quality, the lack of shared prioritization criteria and manual triage.
Based on this diagnosis, the process could be redesigned before any attempt at automation at scale.
The logic was: map, understand, redesign, test. And only then automate.
Before the definitive solution, the team tested the hypothesis that organizing the flow and supporting the analysis could increase recovery capacity. The MVP confirmed the direction, but it also revealed something more important: technology alone would not solve the problem.
The solution needed to combine three elements: a redesigned process, better-quality data and intelligent decision support. That learning is what defined Recuper.AI.
Recuper.AI
Recuper.AI is a virtual assistant used by the IGEPPS legal team to support the analysis of debt recovery cases.
The solution receives case documents, extracts relevant information, identifies debt-related data, locates applicable legal provisions and produces structured justifications that support the drafting of collection letters.
AI does not replace the legal decision.
It reduces the repetitive work that happens before that decision, organizing information, checking criteria and giving the team a clearer view of the cases that deserve attention.
As a result, the legal team no longer works as a checklist executor and can devote more time to work that requires interpretation, judgment and strategy.
From an undifferentiated queue to a manageable portfolio
The most important change was not just speeding up an activity.
It was turning debt recovery into an operation that can be observed and managed.
Previously, triage was manual, cases arrived with different levels of information and there was no single prioritization criterion.
With the new flow, analysis became assisted, information can be structured more consistently and the queue can take into account criteria such as viability, deadlines and case context.
Implementing the new flow changed the way debt recovery is conducted at IGEPPS. The collection portfolio is now monitored in a more structured way, with clearer criteria for analyzing and prioritizing cases, increasing visibility over the flow and supporting more proactive management.
Recuper.AI reduced repetitive triage and document organization tasks, allowing the legal team to focus more time on situations that require interpretation, judgment and strategy.
The solution also expanded the capacity to analyze cases, increased information traceability and reduced the risk of recovery opportunities being lost due to delays or lack of visibility.
More important than automating a task, the project turned debt recovery into a process that can be observed, monitored and continuously improved.
AI in production, with enterprise architecture
Recuper.AI needed to go into production without waiting for the next full modernization of the Institute's infrastructure.
The Red Hat OpenShift environment already in place at IGEPPS served the agency's systems, but it did not have adequate GPU capacity for the new AI workloads.
Waiting for a new infrastructure cycle would have delayed the journey.
Vibe and Red Hat therefore designed a hybrid architecture, able to deliver value in the short term without abandoning IGEPPS's long-term technology strategy.
The solution uses:
- Red Hat OpenShift dedicated to AI workloads;
- cloud infrastructure with GPU;
- Red Hat AI;
- Red Hat AI Inference Server;
- a Qwen model served via vLLM;
- Docling for reading and extracting content from scanned documents;
- an application structured into specialized components for different stages of the process;
- Redis for state management;
- a React frontend and a services layer.
The architecture makes it possible to evolve AI capabilities without creating a solution isolated from the Institute's enterprise ecosystem.
Sensitive data remains under control
In an application that works with legal documents and social security information, performance was not enough.
The architecture had to ensure governance, traceability and control over the data being processed.
Recuper.AI was designed so that the institution keeps control over the infrastructure and over how the models are served, keeping AI within a governed enterprise platform aligned with the IGEPPS technology ecosystem.
Before going into production, the solution went through two levels of institutional validation:
- the business area validated its fit with the process;
- the technology area validated the architecture, platform and integration with the institutional environment.
The final decision remains human.
Built together with those who know the problem
The project was carried out with the Vibe team working directly with the IGEPPS core business areas in Belém.
This proximity shortened the interval between discovery, prototype, validation and product evolution.
The working model started from a simple premise:
those who know the process deeply do not need to learn how to build software; those who build the technology need to learn to listen to those who live the process.
Recuper.AI was born from that encounter.
The first AI in production opened the door to the next ones
Recuper.AI became the first Artificial Intelligence application in production at IGEPPS, but the project's impact went beyond debt recovery.
Other areas of the Institute began evaluating processes that could also be supported by AI.
At the same time, infrastructure modernization and the expansion of institutional Artificial Intelligence capacity became part of the agency's technology evolution.
The discussion is no longer about an isolated application but about an institutional AI capability.
A case presented nationally
The Recuper.AI case, built with IGEPPS, took 1st place in the Equipes category of Premiação Impacto Brasil 2026, by Agile Trends, with Vibe Tecnologia as its partner.
The project came to be presented as a reference for the application of Artificial Intelligence in the public sector.
The case was presented at the Red Hat Gov Forum, in Brasília, in a session with IGEPPS participation, curated by Red Hat.
Later, Vibe Tecnologia and IGEPPS jointly presented Recuper.AI at Agile Trends Gov 2026, in Brasília, showing not only the technology built but, above all, the process transformation that made its application possible.
The experience also received institutional coverage from the Pará State Government.
Next steps: from assistant to recovery copilot
Recuper.AI started by supporting legal analysis and the structuring of collection cases.
The next stage expands that role.
The vision is to evolve the solution into a copilot for the debt recovery management area, connecting data, institutional criteria, intelligent analysis and operational monitoring in a single flow.
Automatic data intake and consolidation
The solution is set to bring together information currently spread across different sources and systems, reducing manual, fragmented lookups.
The goal is to consolidate the evidence and data the process needs into a single operational journey.
Intelligent analysis and viability engine
Documents and evidence are to be read and cross-checked automatically.
The AI engine can support classifying cases into categories such as:
- viable;
- viable with caveats;
- pending additional information;
- unviable.
The final decision remains with the responsible civil servant, who can approve, adjust or request additional information.
Support for collection execution
Beyond analysis, the platform is set to evolve to support later stages of the flow, such as:
- notification;
- handling of appeals;
- setting up payroll deductions;
- operational monitoring.
Recuper.AI would no longer support just one point of the journey and would contribute to the recovery flow as a whole.
Portfolio management and monitoring
The planned evolution also includes a management layer with dashboards and indicators to provide continuous visibility into:
- case progress;
- process bottlenecks;
- productivity;
- information quality;
- recovery progress.
The goal is to move from a reactive operation to data-driven recovery governance.
The principle remains the same
Even in this evolution, AI does not make the decision in place of the civil servant.
It consolidates evidence, applies criteria, suggests paths and reduces repetitive work.
The opinion remains a recommendation.
Validation remains human.
Technology expands the team's capacity; it does not replace its responsibility.
Technology as an enabler. Not as the protagonist.
Recuper.AI was not born from a question like:
“Where can we use artificial intelligence?”
It was born from a more important question:
“Where is the process losing value, and what needs to change to recover it?”
- 01First came Lean Thinking.
- 02Then VSM.
- 03Then the flow redesign.
- 04Then the MVP.
- 05Then better-structured data.
- 06Only then came intelligent automation.
This is how Vibe understands Artificial Intelligence applied to critical processes:
not as a layer placed on top of disorder, but as a capability built on a process we already know how to improve.
Enterprise AI
Models, agents, governance and AI operations in enterprise environments.
Do you have a critical process that still depends on manual work to function?
Vibe helps you understand the flow, redesign the process and build the technology needed to turn operations into institutional capability.