git:20260708.0cbcc56 to git:20260722.5759383

# LIA — Multi-Agent Conversational AI Assistant

> LIA (Learning Intelligent Assistant) is an open-source, multi-agent conversational AI platform

> that orchestrates 19+ specialized AI agents to manage daily productivity — emails, calendar,

> contacts, files, tasks, smart home and more — with systematic human validation (HITL) of every

> sensitive action. Nearly 100% of its code was written by an AI under human direction; the method

## Key Features

- **19+ specialized AI agents**: emails, calendar, contacts, files, tasks, weather, web search, Wikipedia, places, routes, reminders, health metrics, browser automation, smart home (Philips Hue), and more — 76 integrated tools

- **7 LLM providers**: OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, Perplexity, Ollama (local models) — per-node model selection, cost comparison, no vendor lock-in

- **Two execution modes**: a deterministic Pipeline (plan → validate → approve → parallel execution, 4–8× fewer tokens) and an autonomous ReAct loop — user-switchable per conversation

- **Human-in-the-Loop (HITL)**: 6 approval levels ensuring no sensitive action happens without user consent; pending approvals survive restarts

- **Today briefing**: personalized home dashboard — synthesis, weather, agenda, birthdays, reminders, health, unread emails

- **Psyche Engine**: dynamic emotional state — 14 moods, 22 emotions, Big Five personality traits, a relationship that deepens over time ("show, don't tell")

- **Personal journals**: stratified introspection (observations → directives → patterns → portrait) with deferred self-evaluation and a compiled user-model portrait

- **Connected services**: Google, Apple iCloud, and Microsoft 365 (email, calendar, contacts, tasks, Drive) — one active provider per category, interchangeable

- **Voice mode**: wake word and speech-to-text run 100% locally in the browser (Sherpa-onnx/Whisper, 99+ languages); TTS via Edge (free), OpenAI or ElevenLabs

- **Knowledge Spaces (RAG)**: upload documents in 15+ formats or sync Google Drive folders — hybrid semantic + keyword search with per-query cost transparency

- **AI image generation & editing**: text-to-image and photo editing with cost tracking

- **Interactive outputs**: MCP Apps widgets, Excalidraw diagram generation, rich skills rendering maps/dashboards/mini-apps in chat, LaTeX math rendering

- **Browser automation**: autonomous web navigation via Playwright with progressive screenshot streaming

- **Extensible via MCP**: per-user MCP servers (OAuth 2.1, SSRF protection) with an iterative ReAct mode for complex APIs

- **Skills system (agentskills.io)**: create custom expert skills in natural language, import from the marketplace, rich HTML/image outputs

- **Sub-agents**: delegate complex research to specialized ephemeral experts working in parallel

- **Proactive intelligence**: heartbeat notifications (weather changes, meetings, tasks), interest tracking, scheduled recurring actions, reminders

- **Multi-channel**: web app (installable PWA), Telegram (chat, voice, HITL inline buttons), push notifications

- **Health metrics**: heart rate and step ingestion from any source (iPhone Shortcuts documented), period charts, opt-in assistant access

- **Usage quotas**: per-user token/message/cost limits with 5-layer enforcement and real-time gauges

- **14 personalities**: from Professor to Cynic — each user gets their own assistant persona and emotional relationship

- **6 UI languages**: French, English, German, Spanish, Italian, Chinese — parity enforced in CI

- **Cost transparency**: every message shows its token count and cost in euros; per-user CSV exports

## Architecture

- **Backend**: Python 3.12+ / FastAPI with Domain-Driven Design (31 bounded contexts)

- **Frontend**: Next.js 16 / React 19 / TypeScript strict / TailwindCSS

- **Orchestration**: LangGraph 1.x state machine — conditional routing, parallel execution waves, PostgreSQL checkpointing

- **Database**: PostgreSQL 16 + pgvector (state, vectors) + Redis (cache, rate limiting, sessions)

- **Security**: OAuth 2.1 with PKCE, BFF pattern (HTTP-only cookies), Fernet encryption, PII filtering, GDPR compliant

- **Self-hosting**: multi-arch Docker images (amd64/arm64) — runs in production on a Raspberry Pi 5

## How It Works

1. **Ask naturally** — text, voice, or attachments, in any language

2. **LIA plans** — the orchestrator analyzes, identifies agents, builds an execution plan

3. **You approve** — before any sensitive action, LIA shows a preview and waits for consent

4. **LIA executes** — specialized agents run in parallel and deliver real-time results via SSE

## Use Case Examples

- "Find the 3 best cafés near me that are open now and give me walking directions"

- "Send an email to my wife and son to explain I'll be 30 minutes late, use a humorous tone"

- "What time should I leave to arrive at Bastille by 9pm via public transit?"

- "Create a reminder the day before at 7pm for each of my appointments this week"

- "Which day this week is best for an outdoor walk, and suggest 3 nearby places with a nice café"

## Links

- [Landing Page](https://lia.jeyswork.com)

- [Story — how LIA was built by an AI under human direction](https://lia.jeyswork.com/story)

- [FAQ](https://lia.jeyswork.com/faq)

- [Technical deep dive](https://lia.jeyswork.com/how)

- [Philosophy](https://lia.jeyswork.com/why)

- [GitHub](https://github.com/jgouviergmail/LIA-Assistant)