Our recent Medium articles
- Kyohei Yashima
- Nov 14
- 2 min read
Helios recently published two new articles on Medium.
Here is an overview of the latest content we’ve shared.
Dev Tips: How to Get Structured JSON Output with LangChain and Pydantic
In our latest Medium article, we explore a common pain point for developers working with large language models: inconsistent output formatting. If you've ever written fragile string-parsing logic just to extract simple information, this guide shows a better way.
The article introduces a practical method to force LLMs to return predictable, structured JSON every time by combining LangChain with Pydantic. LangChain manages prompt workflows, while Pydantic ensures strict schema validation using Python type hints.
We walk through a three-step process:
Define a data model using Pydantic’s BaseModel.
Generate formatting instructions using PydanticOutputParser, which tells the LLM exactly how to structure its output.
Build a LangChain pipeline that sends the instructions to the LLM and parses the response directly into Python objects.
The article also showcases a practical use case: generating synthetic user personas that always follow a predefined schema. By combining Pydantic models with LangChain prompts, developers can automatically create consistent test data such as user profiles with controlled fields like name, age, hobbies, and job.
Overall, this technique eliminates unpredictable free-form outputs and enables developers to rely on LLM responses as structured, validated data—making applications more robust and easier to maintain.
Measuring the World Beneath the Wheels
In this Medium article, we explore how an everyday smartphone can function as a compact environmental sensor. What began as a small internal prototype at Helios developed into a Kotlin-based Android app capable of recording road inclination using accelerometers, gyroscopes, magnetometers, and GNSS data.
The app remains lightweight and fully local, offering angle correction, stationary detection for recalibration, continuous CSV logging, and direct compatibility with QGIS for spatial analysis. Through field tests ranging from short rides to long-distance driving, the system captured dense elevation traces that corresponded closely with DEM terrain models, while also revealing challenges such as acceleration bias and the need for improved filtering.
The article highlights key engineering considerations—Android permission changes, Gradle and JDK adjustments, sensor-fusion refinements—and reflects on our rapid prototyping workflow using GitHub and CodeX. We conclude with next steps, including GNSS–IMU synchronization, real-time visualization, and potential deployment in northern Okinawa for regional monitoring initiatives.
Comments