AI & Data Transparency

Effective Date: May 2026

Contents

AI Transparency

1. Overview

2. Context Service

3. Operator Service

4. Integrations and data use

Data Transparency

1. Overview

2. Infrastructure, hosting and operations

3. AI processing and inference

4. Communication, voice and telephony

5. Product operations and internal systems

6. Provider legal information

This page provides information about the AI and ML components that underpin the dConnection FreeMe service, as well as the compute, data and service providers used to deliver and operate it.

AI Transparency

1. Overview

FreeMe uses different AI and ML components to understand conversations, prioritise content, process voice, support search and enable responses within the product. Depending on the function, models are operated through OpenRouter in Zero Data Retention mode, through native provider integrations or on dConnection self hosted infrastructure.

2. Context Service

The Context Service helps FreeMe structure, summarise, retrieve and maintain information from connected channels so that it can be used as part of the user context.

2.1 Summaries and core memory. For messenger conversation summarisation, agent and SMS conversation summarisation, email thread summarisation, conversation segment merge decisions, core memory extraction, core memory update and deduplication decisions, as well as voicemail summarisation as a fallback, dConnection uses Google Vertex AI through OpenRouter. The model used is google/gemini-3.1-flash-lite.

2.2 Urgency classification. For external content urgency classification from Gmail and Google Calendar source channels, dConnection uses Google AI Studio through a native integration. The model used is google/gemini-3.1-flash-lite.

For external content urgency classification from other channels, dConnection uses Google Vertex AI through OpenRouter. The model used is google/gemini-3.1-flash-lite.

2.3 RAG, search and relevance ranking. For RAG intent detection, synthetic query generation, contact name extraction and the LLM reranker fallback path, dConnection uses Cerebras, Groq, Google Vertex AI, BaseTen and DeepInfra through OpenRouter. The model used is openai/gpt-oss-120b.

For the RAG fallback path, dConnection uses DeepInfra, Nebius and SiliconFlow through OpenRouter. The model used is qwen/qwen3-embedding-8b.

2.4 Self hosted reranking and embedding models. For reranking, dConnection uses self hosted infrastructure through a native integration. The model used is qwen/qwen3-Reranker-4B.

For RAG and core memory query embeddings, email thread and message chunk indexing, messenger segment and summary indexing, agent and SMS segment and summary indexing, voicemail transcript and summary indexing, core memory document embeddings, as well as transient messenger and transient agent storage, dConnection uses self hosted infrastructure. The model used is qwen/qwen3-Embedding-8B.

3. Operator Service

The Operator Service supports active interaction with FreeMe. This includes chat, tool calling, AI Voicemail, speech recognition, speech synthesis and selected post processing tasks.

3.1 Main chat and tool calling. For the main chat conversation and tool calling, dConnection uses Google Vertex AI and Amazon Bedrock through OpenRouter. The model used is anthropic/claude-sonnet-4.6.

3.2 AI Voicemail. For AI Voicemail conversations, tool calling and AI Voicemail call summarisation, dConnection uses Google Vertex AI and Amazon Bedrock through OpenRouter. The model used is anthropic/claude-haiku-4.5.

3.3 Agent and agenda processing. For generating the opening line in agent calls, dConnection uses Google Vertex AI through OpenRouter. The model used is google/gemini-3.1-flash-lite.

For agenda post processing, including the detection of whether content was presented or interrupted, dConnection uses Google Vertex AI through OpenRouter. The model used is google/gemini-3.1-flash-lite.

3.4 Voice input and voice output. For user voice transcription, also known as speech to text, dConnection uses Soniox through a native integration. The model used is Soniox/stt-rt-v4.

For assistant voice synthesis, also known as text to speech, dConnection uses ElevenLabs in the EU through a native integration. The model used is eleven_flash_v2_5.

3.5 Load testing. For main chat conversations and tool calling in load testing mode, dConnection uses Venice and DeepInfra through OpenRouter. The model used is google/gemma-4-31b-it.

4. Integrations and data use

OpenRouter is used exclusively with Zero Data Retention mode enabled. Your data is never used to create, train or improve a machine learning or artificial intelligence model. Your data is never stored by either OpenRouter or our inference providers.

Native integrations use the provider's first party tooling, API, SDK and infrastructure without intermediaries. For self hosted models, dConnection directly uses vLLM. For provider hosted models, dConnection uses the provider's first party cloud through that provider's SDK or API.

Data Transparency

1. Overview

The following sections list the compute, data and service providers used by dConnection for infrastructure, payments, communication, error reporting, localisation, source control, voice, AI inference and internal systems.