Pratibmb

The problem

The most intimate dataset you own is a decade of private messages — and it's exactly the dataset you should never upload to somebody else's API. I wanted to build something meaningful on top of that data without it ever leaving the machine it lives on.

What I built

Pratibmb ingests your personal messaging history, builds a private corpus on your computer, and lets you have a conversation with the person who sent those messages years ago. Pick a year on a slider and start texting past-you.

It imports from eight platforms — WhatsApp, Facebook, Instagram, Gmail, iMessage, Telegram, Twitter/X, and Discord — parsing each export format into a local SQLite database. A GGUF embedding model (84MB) runs through llama.cpp to embed the corpus. A local LLM (2.3GB) then profiles the conversations to extract relationships, life events, interests, and communication style, so replies are grounded in real memories rather than invented ones. Optional LoRA fine-tuning teaches the model your specific texting habits — abbreviations, language mix, emoji patterns.

Interesting decisions

• One non-negotiable constraint: it must run with the Wi-Fi off. No cloud, no telemetry, no API calls. Every design decision downstream falls out of that single rule.
• Local models sized to fit real machines. An 84MB embedder and a 2.3GB chat model, rather than something that only runs on a workstation.
• AGPLv3. For software that touches this kind of data, the licence should guarantee anyone can verify what it does.
• Ship it like real software. Packaged for macOS, Windows, and Linux, and distributed through Homebrew, the AUR, and winget — not just a zip file.

Outcome

Live at pratibmb.com with downloads for all three desktop platforms. The name is a coined word — a distinctive mark for a distinctive idea.