AADITHYA VIMAL

AI · 2026 · Live demo · Open source · 2026

Pith Analytics

A local-first, privacy-focused data analytics platform running DuckDB WebAssembly directly inside the browser with on-device WebLLM natural language SQL translation.

My role
Designed and built the in-browser OLAP analytics workspace, WebLLM integration, and Vega-Lite visualization pipeline.
Status
Live demo · Open source · 2026
Category
AI · Local browser analytics platform
Live demoSource on GitHub

Live demo · interactive

🔒pith-analytics.pages.dev
CONNECTING TO PITH ANALYTICS...

Embedded for convenience — if it doesn’t load here, open the full site directly.

01

Understand

Why it exists. Analyzing sensitive spreadsheets or databases traditionally requires uploading confidential files to remote servers. Pith brings full SQL analytical power entirely into the browser.

What I built. Local-first, browser-based data analytics suite executing high-performance analytical queries directly on the client machine using DuckDB compiled to WebAssembly, paired with on-device natural language AI assistance.

The problem it solves. Querying and visualizing sensitive or confidential datasets typically forces organizations to upload confidential data to third-party cloud data warehouses, creating privacy liabilities, compliance headaches, and recurring infrastructure costs.

What it does

  • DuckDB compiled to WebAssembly executes SQL across millions of rows with near-native performance
  • 6 on-device AI models via WebLLM/WebGPU (Llama 3.2, Qwen 2.5, Phi 3.5) translating plain English to SQL
  • Multi-format ingestion supporting drag-and-drop CSV, JSON, and Parquet files
  • Interactive data visualizations powered by Vega-Lite with zero server roundtrips or telemetry
02

Technical depth

Architecture. Ingests CSV, JSON, and Parquet files into in-browser DuckDB WASM in-memory tables capable of scanning millions of rows in milliseconds. Natural language queries are translated to executable SQL locally using 6 selectable WebLLM / WebGPU models (Llama 3.2, Qwen 2.5, Phi 3.5), with instant interactive chart rendering via Vega-Lite.

Security model. 100% private and offline-capable: zero network requests, zero telemetry, and zero remote logging. Data never leaves client device memory and persists solely in local IndexedDB. Fully open-source under AGPL-3.0.

Tradeoff. Execution is bounded by available client device RAM and GPU capabilities, prioritizing absolute data privacy over multi-node cluster scaling.

  • React
  • TypeScript
  • DuckDB WASM
  • WebGPU
  • WebLLM
  • Mosaic
How it works, step by step

Ingests CSV, JSON, and Parquet files into in-browser DuckDB WASM in-memory tables capable of scanning millions of rows in milliseconds. Natural language queries are translated to executable SQL locally using 6 selectable WebLLM / WebGPU models (Llama 3.2, Qwen 2.5, Phi 3.5), with instant interactive chart rendering via Vega-Lite.

Security notes

100% private and offline-capable: zero network requests, zero telemetry, and zero remote logging. Data never leaves client device memory and persists solely in local IndexedDB. Fully open-source under AGPL-3.0.

AI