How Tendy turned LLMs into a Business OS with memory, RAG and integrations

Tendy was born from a simple thesis: companies should not depend only on generic chatbots to use artificial intelligence. The real value of an LLM appears when it understands company context, accesses internal data, executes actions in business systems, keeps long-term memory, and can be distributed through an app, API, and partner channels. The challenge was turning that vision into an operational AI Business OS.

The architecture was designed as a multi-tenant platform with an authenticated app, Cloudflare Workers API gateway, persistent conversation history, shared projects, document upload, RAG, integrations, billing, and model controls. Instead of coupling everything to a chat screen, the LLM logic was treated as infrastructure: routing, context, tools, billing, and distribution need to work together.

LLM gateway with aliases and routing

At the core of the product, I structured a gateway compatible with the chat completions pattern. The platform exposes public model aliases, such as fast or more analytical assistants, while keeping raw provider model IDs private. This makes it possible to change internal routes without breaking the customer experience or the public API.

I also implemented classification and routing logic so an automatic model can delegate the response to the most appropriate alias. This separation between public alias, provider model, and assistant behavior made the product more flexible: the user chooses intent, without needing to understand the infrastructure complexity behind it.

RAG, memory and proprietary context

The second layer was context. Tendy separates knowledge by tenant and RAG profile, using document ingestion, R2, queues, and Vectorize to build a retrieval layer. The goal was to let each company connect its own documents and data without mixing context across customers.

Beyond RAG, the platform includes a memory layer with providers such as Mem0 and Zep. This memory was designed to store preferences, persistent facts, and useful workspace history. The distinction matters: RAG retrieves documentary knowledge; memory helps the LLM remember operational context and recurring company preferences.

Tool calling and business integrations

A corporate LLM cannot only answer. It needs to execute. For that reason, Tendy integrates hundreds of systems through Nango and turns those systems' actions into OpenAI-compatible tool calling functions. The platform normalizes tool formats, filters executable actions, and injects useful descriptions so the model can decide when to use each tool.

This design lets the assistant work connected to business tools without exposing technical complexity to the user. The result is no longer an isolated conversation, but an operations interface: the LLM can query, trigger, and organize workflows in external systems with tenant-level governance.

Dedicated deployments, billing and distribution

I also structured the path for dedicated GPU models using RunPod, with tenant-level deployment creation, model aliases, and simulation through Cloudflare Workers AI when needed. This layer prepares the product for scenarios where a company wants more control over model, capacity, and infrastructure.

The operation was completed with measurement and monetization: app subscriptions, direct API usage, token metering, Stripe, and distribution through partner channels or marketplaces. That turned Tendy into more than an AI application. It became commercial LLM infrastructure: team app, developer API, memory, RAG, integrations, dedicated deployments, and billing in the same system.