Lidless Oracle
jimyeong/lidless-oracle
Chatbot de risco de mofo com sensores IoT e Claude multilíngue
Ver no GitHubChatbot multilíngue integrado com Claude e dados reais de sensores IoT para avaliação de risco de mofo em residências do Reino Unido. Fornece aconselhamento conversacional e personalizado sobre prevenção e mitigação de mofo, com suporte a múltiplos idiomas e cálculos baseados na fórmula de Magnus.
Instalação
npx convez-ai jimyeong/lidless-oracle1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Avaliação de risco de mofo em habitações usando dados de temperatura e umidade em tempo real
- Suporte multilíngue para comunidades diversas em habitação social
- Integração de dados de sensores Zigbee com raciocínio de IA para aconselhamento contextualizado
Veredito

Projeto robusto que exemplifica bem a orquestração de Claude com dados IoT reais e múltiplas linguagens. A arquitetura multi-serviço (Node + Python) é sólida, embora ainda em desenvolvimento ativo com funcionalidades como persistência de conversas e alertas em tempo real por vir.
README
A tradução para o português deste README ainda está na fila.
Lidless Oracle 🔮
A multilingual LLM chatbot for mould risk management in UK homes.
What this project is
Mould affects ~4% of UK homes (English Housing Survey) and disproportionately impacts social housing tenants and non-native English speakers — those who most need accessible information often face the steepest barriers to it.
Lidless Oracle is a chatbot designed to bridge that gap. Users converse in their own language about mould — its causes, prevention, and the present risk in their home — while the system fetches real environmental data from IoT sensors and reasons over it through an LLM (Claude).
Why a chatbot, why multilingual
- Conversational over forms: most mould advice today comes as static guides or symptom-checking forms. People don't speak in form fields.
- Multilingual by design: UK housing reaches diverse communities. An advisor that answers in Korean, Polish, Arabic, or Urdu meets users where they are. The LLM handles this natively, without a translation layer.
- Grounded in real data: rather than generic advice, the system queries actual sensor readings (humidity, temperature) and assesses risk via the Magnus formula in a separate Python service.
Demo (May 2026)
The current build demonstrates the core conversation loop:
- User asks in natural language: "What was last week's mould risk?" / "이번 주 곰팡이 위험은?"
- Claude interprets the time range, calls a backend tool that fetches real IoT data, and replies conversationally with risk score, sample size, and actionable advice.
Architecture
| Component | Tech | Role |
|---|---|---|
| Mobile client | React Native, TypeScript, Apollo Client | Chat UI, auth, GraphQL |
| API gateway | Node.js, Apollo Server, Prisma | LLM orchestration, tool routing |
| LLM | Anthropic Claude (tool_use) | Natural language understanding, multilingual response |
| Mould risk engine | Python | Magnus formula on Zigbee sensor data |
| Auth | Firebase Authentication + Admin SDK | Google Sign-In, ID token verification |
| Database | PostgreSQL | Users, conversations, messages |
Status & roadmap
🚧 Active personal development.
Done
- ✅ Google authentication via Firebase
- ✅ Conversational chat UI
- ✅ LLM tool use loop (Claude ↔ mould risk service)
- ✅ Real IoT data integration (4,000+ sensor readings)
Next
- ⏳ Conversation persistence per user (separate sessions)
- ⏳ RAG layer with UK housing health guidance (NHS, HHSRS Cat 1 hazards)
- ⏳ Real-time alerts when risk crosses thresholds
- ⏳ iOS support
Notable design decisions
- Multilingual via LLM, not translation layer: Claude responds in the user's language without an explicit i18n system. The system prompt sets language-matching behaviour; output quality is the LLM's responsibility.
- Firebase Auth + own DB: Firebase handles OAuth and token lifecycle; user-domain data lives in Postgres via
firebase_uidforeign reference. Backend verifies tokens with Firebase Admin SDK on every request. - Tool use loop: The chat resolver maintains a loop that handles arbitrary numbers of tool calls per turn, with an iteration limit for safety.
- Multi-service backend: LLM orchestration (Node) is separate from sensor data processing (Python) — each in the language best suited to its job.
Repo scope
This repository contains the React Native client. The backend (lidless-hermes) and IoT subsystem (lidless-controller) are private repositories.
Built solo, May 2026.
