Datrix: Chat With Your Data Using Gemma 4 — Charts, ML Models, No Code

This is a submission for the Gemma 4 Challenge: Build with Gemma 4 Datrix is a chat interface for your data — upload a dataset, ask questions, get charts and trained ML models back. No code required. The problem it solves is simple: most people who work with data aren't developers. They know what they want to find out, they just can't write the code to find it. Datrix bridges that gap by letting Gemma 4 handle the Python while the user just asks questions. Interactive charts — ask for a visualization in plain English and get a rendered, interactive chart back in the browser ML model training — request a trained model, and it gets saved to your session for follow-up predictions in the same conversation Image input — attach a screenshot of a report or a photo of a whiteboard and ask questions about it alongside your data Local or cloud — run fully offline with Gemma 4 E4B on Ollama (~9.6 GB RAM, no GPU), or use the 31B model via OpenRouter for heavier analysis Supported formats: CSV, Excel, JSON, Parquet, and images up to 200 MB. datrix-gemma GitHub Repository 256K context window Data analysis is multi-turn by nature — you explore, refine, go back. The 256K context window means Datrix can include the full dataset schema, session charts, and conversation history in every request. Gemma 4 always has the full picture. Code generation and self-correction Every query becomes a Python script. Gemma 4 writes it, the sandbox runs it, and if it fails the traceback goes back to Gemma 4 to fix automatically. Vision No extra models or OCR setup needed. Images go straight into the Gemma 4 request and get handled natively alongside the rest of the conversation. Local mode and sensitive data A lot of datasets people might actually want to analyze — HR records, financials, patient data, internal metrics — can't go to a cloud API. With Gemma 4 E4B running locally via Ollama, everything stays on your machine. Same experience, fully private, no GPU required. For users who want stronger reasoning and can't fit the 31B model on their hardware, OpenRouter is the easy route. But if your machine can handle it, you can also pull gemma4:31b via Ollama and run it fully locally too (requires ~20 GB RAM at 4-bit quantization, 32 GB recommended for comfortable use) — no data leaving your device, full reasoning power.