AI that works for you. Not against you.

Every incoming event is evaluated against your rules in under 100ms. 7 synchronous operators (person groups, time ranges, channel types, named locations, active call status, online presence, device connectivity) plus 4 async operators that query contact timelines (call count, last contact date, unread voicemails, missed calls). 11 action types fire automatically — and any action supports delayed execution via schedule.

EmailAIService processes every incoming email automatically. Thread summarization extracts a summary, action items, and key points from the full conversation. Reply suggestions generate three tones: professional, friendly, and brief. Priority scoring rates threads 0-100 based on urgency keywords, deadline mentions, sender importance, and direct-vs-CC addressing. Follow-up detection scans your outbound messages for commitments and suggests due dates.

TranscriptionService subscribes to CallEnded events and triggers a six-step pipeline for calls over 10 seconds: fetch recording from Twilio, transcribe via Whisper STT, extract a summary and sentiment analysis, identify action items and key topics, pull named entities, and feed everything into the LightRAG knowledge graph. Results are stored in call_transcripts and call_analyses with full segment-level detail.

pgvector embeddings are generated automatically for every new message, document, and event via the EmbeddingPipelineService. Five Meilisearch indexes (persons, messages, events, documents, listings) provide instant full-text search with 30-second result caching and circuit breaker degradation. Natural language queries hit both systems for hybrid ranking.

EmbeddingPipelineService subscribes to EntityCreated and EntityUpdated events to generate embeddings at write time using configurable models (OpenRouter or Ollama). LightRAGBridgeService extracts entity relationships from every piece of content and builds a queryable knowledge graph. WikiService compiles persistent knowledge pages from event-driven auto-ingest.

ToolRegistry provides 28 tools the AI agent can invoke, including search_contacts, get_contact, search_messages, list_rules, search_documents, analyze_document, search_entities, process_file, list_tool_capabilities, search_wiki, update_wiki, get_contact_timeline, search_comms, get_unread_count, get_call_summary, send_message, query_observations, log_observation, get_current_context, find_correlations, list_intentions, get_habit_status, and get_health_summary. Tools are conditionally registered based on available services.

MultiAgentOrchestrator decomposes complex goals into step-by-step plans using Claude 3.5 Haiku. Each step can invoke a registered tool or perform reasoning. Steps execute sequentially with full error handling — the plan continues even if individual steps fail. AgentMemoryService provides working memory with optional TTL and complete tool call logging for debugging.

PolicyService auto-redacts phone numbers, email addresses, and SSNs before any text reaches a cloud LLM. Consent checking gates AI processing, federation sharing, and location tracking individually. All LLM interactions are audit-logged with provider, model, and routing decision. Run everything locally with Ollama — or route tasks to different models based on sensitivity and complexity.