Batchwise

• 

  Multi-language translation with specific tones

• 

  News headline categorization

• 

  Car make & model extraction

Inspiration

I developed Batchwise to streamline my data processing workflow with LLMs. While I frequently used API calls to services like OpenAI, Anthropic, and Google Cloud for data processing, I found myself repeating similar code patterns. Moreover, concerns about data privacy and API costs led me to seek a solution that could process data locally. Batchwise emerged as a no-code tool that addresses these three key needs: operational efficiency, data privacy, and cost-effectiveness.

What it does

Batchwise is a no-code tool that processes CSV data using natural language instructions. It leverages Chrome's built-in AI capabilities to transform data row by row, all while keeping processing local to the user's device.

How I built it

The application is built as a React web application, developed with assistance from Github Copilot. The implementation relies on open-source libraries for CSV handling and includes an Excel-like spreadsheet interface for user interaction.

Challenges I ran into

Working with Chrome's AI API presented several technical challenges. The API's instability led to occasional NotReadableError and NotSupportedError exceptions. Additionally, the API lacks a deterministic way to generate tokens which sometimes impedes debugging.

Accomplishments that I'm proud of

Batchwise is arguably more useful than my previous Chrome AI experiment (https://github.com/zmxv/fecal).

What I learned

Fine-tuned Gemini Nano models optimized for translation and rewriting tasks outperform the general-purpose model, even with prompt engineering tricks.

What's next for Batchwise

A hybrid architecture that combines the benefits of both local and remote LLMs, offering users flexibility in choosing their data processing approach for specific needs.

Built With

• react
• typescript