Coomecipar - Conversational Assistant for Document and Knowledge Management

As part of the Pro Internacional (prointernacional.com) and Delfos (delfoslabs.com) teams, I contributed to the development of a private conversational assistant platform for Coomecipar (coomecipar.coop.py), a major financial cooperative in Paraguay. The goal was to give the organization the benefits of LLM-powered assistance without exposing internal data to public chat tools or losing control over privacy and access.

Coomecipar needed a way for employees to ask questions about internal processes and documents in natural language, but with strict constraints around data privacy, security and accuracy. Using public LLM interfaces was not an option for them, so the decision was to build an internal-only chat platform, fully integrated with their infrastructure and identity systems.

The application layer was built with:

• Payload CMS as the backbone for managing content and organizational knowledge used by the assistant.
• Next.js as the frontend framework, using shadcn/ui to build a clean, accessible chat interface and admin tools.
• A MongoDB database as the main persistence layer for conversations, user context and knowledge-related metadata.

For the AI layer, we implemented:

• A FastAPI backend in Python acting as the interface to Anthropic models, consumed through Microsoft's cloud services.
• An integration with the organization's internal SSO based on Active Directory, so only authenticated employees can access the assistant and all usage is tied to corporate identities.
• A minimum but curated internal knowledge base that the assistant can use as context, helping it stay grounded in Coomecipar's language, policies and processes instead of generic internet knowledge.

The result is a fully private, enterprise-focused chat platform where:

• Users log in via internal SSO (Active Directory).
• Interactions with the LLM stay within the organization's controlled environment.
• The assistant can leverage a knowledge base aligned with Coomecipar's documentation, helping staff get faster, more consistent answers while respecting the institution's privacy and compliance requirements.

My contribution focused on the full-stack implementation of this solution: integrating Payload CMS, Next.js + shadcn and MongoDB on the application side, wiring FastAPI to Anthropic's models via Microsoft, and helping shape the authentication and knowledge flows so the assistant is both useful and safe for a highly regulated financial environment.