Tempestas

Tempestas AI transforms raw IoT telemetry into actionable foresight. By analyzing 60-minute sensor windows with Gemini 2.5, it predicts environmental shifts an hour before they occur.

Comment

Tempestas AI was born from the challenge of making sense of the "noise" in environmental IoT data. While standard sensors provide a snapshot of the present, they rarely offer a glimpse into the near future without complex, hard-coded statistical models.

This project was inspired by the need for a unified ecosystem where data flows seamlessly from hardware to cloud, and eventually to the user's pocket. My goal was to see if Large Language Models (LLMs) could act as zero-shot predictors for time-series data, treating sensor trends like a narrative that the AI can "complete."

How We Built It

We architected this project as a full-stack distributed system:

The Core: A FastAPI microservice handles the heavy lifting of AI inference, using Gemini 2.0 to process 60-minute windows of telemetry.

The Backbone: A C# .NET Web API manages the business logic, device authentication, and coordinates data flow between the sensors and the AI.

The Storage: PostgreSQL (via Supabase) serves as our "source of truth," storing historical records and device metadata.

The Interface: A React dashboard provides real-time visualization, while a React Native (Expo) app ensures users can monitor their environment on the go.

Challenges Faced

One of the primary hurdles was maintaining type safety across three different languages (C#, Python, and TypeScript). Ensuring that a sensor record sent from C# was correctly validated by Pydantic in Python and then rendered accurately in React required a strict shared interface.

Additionally, ensuring the AI understood the "physics" of the data was critical. To help the model, we implemented a windowing logic where the prompt includes the current time t and asks for a prediction at t+60 min.

What We Learned

We mastered the art of Structured AI Outputs. We also learned that instead of asking an AI to "return JSON," it is far more effective to provide a rigid JSON Schema. This ensures that the output y always maps to the expected interface I, such that: y∈{x∣valid_schema(x,I)}

Tech Stack

Backend & Services

C# .NET 8: The primary backend handling business logic, API orchestration, and database interaction.

FastAPI (Python): Dedicated AI microservice for Gemini 2.0 and Ollama integration.

PostgreSQL: Relational database for persistent storage of telemetry and device logs.

Frontend & Mobile

React: Web dashboard featuring interactive charts and real-time data streaming.

React Native & Expo: Cross-platform mobile application for iOS and Android, providing push notifications for environmental alerts.

Artificial Intelligence

Google Gemini 2.5 Flash: High-speed cloud-based time-series forecasting.

Ollama (Phi-4-Mini): Privacy-focused local inference for offline or edge scenarios.

Deployment

Render: Cloud hosting for the FastAPI and .NET services.

Supabase: Managed PostgreSQL and Realtime engine.
Share this project:

Updates