Engineering Agent Hub: How UPIpe Built an Internal AI Platform

Over the last few years, UPIpe has been exploring practical ways to bring Artificial Intelligence into engineering, operations, and internal business processes. Our initial approach involved evaluating external AI platforms that could automate complex, knowledge-intensive workflows. However, as our requirements became more specialized, we discovered that many third-party solutions could not provide the scalability, security, integrations, and flexibility our teams needed.

This led us to develop our own internal application layer, Agent Hub—a production-focused AI platform designed around UPIpe's engineering and operational requirements.

Why External AI Platforms Were Not Enough

Our primary goal was straightforward: use AI to reduce manual effort and improve the efficiency of internal processes. While several commercially available AI platforms offered useful capabilities, we repeatedly encountered limitations when applying them to large-scale enterprise workflows.

Standard AI development environments often came with request restrictions, document-processing limits, storage constraints, and context-window limitations. These challenges became particularly significant for engineering teams working with large internal repositories and extensive knowledge sources.

Large-scale technical integrations also introduced additional complexity. Processing substantial code repositories, internal documentation, and operational data required infrastructure that could handle high volumes without compromising performance or security.

UPIpe also operates in an environment where data governance, privacy, and regulatory requirements are important considerations. Certain workflows require information to be collected from official sources and processed reliably. Conventional web-scraping solutions often lacked the accuracy, consistency, and operational controls required for these use cases.

Another challenge was the need for human-in-the-loop workflows. Some AI processes cannot be completed entirely autonomously and require employees to review information, approve actions, or continue the process through existing operational systems. Moving repeatedly between AI platforms and internal consoles created unnecessary friction.

Rather than adapting critical workflows around the limitations of external products, UPIpe decided to build an application layer that could be designed around its own infrastructure and operational requirements.

Agent Hub: Built for Security, Scale, and Integration

Agent Hub was designed as a specialized internal AI environment capable of supporting production workloads while integrating deeply with existing systems.

Production-Grade Sandbox Execution

Agent Hub uses temporary Docker-based execution environments to isolate workloads. These environments are designed with strict controls, including restricted network access and process limitations.

The platform also works with UPIpe's internal orchestration infrastructure to manage CPU and resource allocation. This allows individual AI workflows to receive controlled compute resources while maintaining predictable infrastructure behavior.

Air-Gapped MCP Architecture

Security and data privacy are central to the platform's architecture. Agent Hub can launch local MCP servers as subprocesses using packages available through UPIpe's internal Artifactory infrastructure.

This architecture minimizes reliance on the public internet and allows tools to operate within controlled internal environments. As a result, teams can build AI workflows while maintaining stronger control over sensitive enterprise data and dependencies.

Deep System Integrations and Bulk Workflows

Agent Hub is designed to work with the information employees already use every day.

The platform can parse attachments such as PDF, DOCX, and CSV files from incoming email conversations and resolve relevant Google Drive references. Instead of requiring employees to process files individually, workflows can automatically execute across large datasets.

For example, a single workflow can process thousands of CSV records or work through files contained in an entire Drive folder. Parent-child workflow tracking provides visibility into individual executions while keeping them connected to the original task.

Unified Semantic Discovery

Finding the right AI agent, tool, or knowledge source can become difficult as an internal AI ecosystem grows.

To address this, Agent Hub uses a common semantic discovery layer. Agents, knowledge bases, and tools are indexed within a unified vector-based collection.

LLM-generated summaries improve discoverability, allowing employees to search for relevant resources using natural-language descriptions rather than remembering exact names or technical identifiers.

Flexible Reasoning and Controlled Tool Management

Agent Hub also provides configurable reasoning capabilities. Step-by-step MCP-based reasoning can be enabled as a tool when an agent requires it instead of being permanently included in every workflow.

The platform includes administrative capabilities for creating, reviewing, curating, and managing tools. Dependency tracking helps teams understand how tools are being used and reduces the risk of unintentionally disrupting agents that depend on production tools.

Built-In ML-Based Anomaly Detection

Operational teams often need to analyze large volumes of monitoring data. Agent Hub addresses this by providing native access to Grafana-based metrics.

An agent can retrieve time-series information from dashboards and perform anomaly detection as part of the workflow itself. This removes several manual steps involved in collecting, interpreting, and analyzing service-level metrics.

A Flexible Multi-Model Strategy

One of the key advantages of owning the application layer is the ability to choose the appropriate model for each task.

UPIpe does not need to rely on a single AI vendor or model family for every workflow. Routine document-processing and classification workloads can be handled using open-source models hosted on internal infrastructure, helping control operational costs.

For workflows that require advanced reasoning or more sophisticated capabilities, the platform can utilize frontier models where appropriate.

This mixed-model architecture provides greater flexibility while reducing unnecessary vendor dependency.

The Impact of Agent Hub at UPIpe

By investing in the workflow and application layer instead of focusing exclusively on models, UPIpe has been able to apply AI across multiple internal functions.

Agent Hub now supports a growing ecosystem of internal AI agents designed to automate repetitive tasks, accelerate information retrieval, and assist teams with complex operational workflows.

The platform currently supports more than 200 internal agents, handling over 1,600 queries every day for approximately 759 Weekly Active Users. Automated tasks save an average of around five minutes per task, creating measurable time savings across teams.

Faster Product Readiness Processes

AI-assisted workflows have significantly reduced the time required for certain internal product-readiness assessments.

Processes that previously required several weeks can now be completed in a matter of days, representing a substantial reduction in turnaround time.

Automating Compliance Reviews

In compliance operations, a lending reverse-audit workflow can analyze loan documentation and identify potential violations automatically.

Tasks that previously required teams to spend several days reviewing documents can now be completed within minutes, allowing employees to focus their attention on cases requiring deeper investigation or human judgment.

Supporting Fraud and Risk Investigations

AI has also become an important part of fraud and risk operations.

Agent Hub supports investigations involving multiple types of information and enables teams to process complex cases much faster than traditional manual workflows.

By using internally hosted open-source models for suitable workloads, UPIpe can also reduce the cost associated with individual investigations. Certain investigations that previously cost around ₹50 per case can be processed at approximately ₹5–₹15, depending on the workflow and model used.

The system has also helped address cases that previously could not receive sufficient human attention because of limited operational bandwidth.

Building AI Around Real Business Workflows

The experience of building Agent Hub demonstrates that enterprise AI is not simply about selecting a powerful model.

For UPIpe, the larger challenge involved connecting AI with real operational systems, controlling infrastructure, managing data securely, supporting human intervention, and creating workflows that could operate reliably at scale.

By developing its own application layer, UPIpe has been able to create an environment where different models, tools, knowledge sources, and internal systems can work together.

Agent Hub provides a foundation for expanding AI adoption while maintaining control over infrastructure, security, integrations, and operational costs.

The result is a production-oriented internal AI platform that helps UPIpe automate repetitive work, accelerate complex processes, improve access to organizational knowledge, and create more efficient workflows across the company.