Swisper¶

Europe's agentic AI platform. Swisper connects people and services through trusted intelligent assistants that plan, coordinate, and execute real-world actions on your behalf.

Code repository: Fintama/helvetiq

What Swisper Does¶

Swisper is a modular AI assistant platform built on three convictions:

Agents are the new interface. People interact with services through intelligent assistants, not apps and websites.
Privacy is non-negotiable. Swiss-hosted, European-grade data protection.
Execution, not conversation. Swisper agents book, modify, transact, and act across connected services.

Platform Modules¶

Core Orchestration¶

Module	Purpose
Global Supervisor	Central orchestrator — routes every conversation through intent classification, memory, agent execution, and response generation
Intent Classification	Determines user intent — simple question or complex multi-step task

Memory and Knowledge¶

Module	Purpose
Fact System	Extracts, stores, and retrieves facts and preferences using PostgreSQL + pgvector
Entity Disambiguation	Resolves ambiguity when multiple entities match a reference
Summarization	Compresses conversations to stay within token limits while preserving context

User Interaction¶

Module	Purpose
UI Response System	Assembles final responses, handles streaming, manages conversation display
Voice System	Speech-to-text and text-to-speech via Azure Speech Services
Greeting System	Personalized greetings based on time, history, and preferences
HITL System	Human-in-the-loop — pauses execution for user clarification

Data & Integrations¶

Module	Purpose
Document Intelligence & RAG	Document upload, indexing, chunking, and retrieval-augmented generation
Integrations	Gmail, Office 365, Telegram, Threema, WealthOS — unified integration management
Background Jobs	Scheduled jobs for email/calendar ingestion, notifications, daily briefings, fact decay
Signals & Notifications	Proactive notification delivery via Telegram and Threema channels

Administration & Operations¶

Module	Purpose
Authentication	Two-factor auth (TOTP), JWT tokens, session handling
Admin Settings	User roles, system parameters, LLM provider selection
Token Usage & Analytics	LLM cost tracking, per-user usage analytics, per-node breakdown
Rate Limiting	Endpoint and token rate limiting for cost control and abuse prevention

Guides¶

Stop Processing — Redis-based stop flag pattern for cancelling in-flight requests
Provider Configuration — Per-node LLM provider configuration
Prompt Asset Pattern — How prompts are managed as .md asset files
Deployment & Testing — Deployment and testing procedures

Decisions¶

Technology Stack¶

Layer	Technologies
Backend	Python 3.12, FastAPI, LangGraph, LangChain
LLM Providers	Google Gemini, Anthropic Claude, Kvant, Azure OpenAI
Database	PostgreSQL + pgvector, Redis
Frontend	React, TypeScript, Vite
Voice	Azure Speech Services (STT + TTS), WebSocket streaming
Infrastructure	Docker, Kubernetes, Helm