Checking backend…
An AI assistant and LLM laboratory, built from first principles.
Libra is a ChatGPT-like assistant with an educational transformer lab underneath — every component of the stack, from tokenization to attention to training, is explainable and inspectable.
Chat Assistant
A ChatGPT-like assistant that answers through configured real model providers — or runs entirely on-device in the educational lab.
Transformer Lab
Learn attention, tokens, training loops and checkpoints by building a tiny transformer from scratch inside the notebook.
Arena & Evaluation
Head-to-head model comparisons, Elo leaderboards, perplexity and latency metrics for every provider in the registry.
RAG Knowledge Base
Upload documents, chunk them, and retrieve answers with hybrid dense + BM25 retrieval and reranking.
Stateful Data Notebook
In-browser analytics kernels with charts, tables and persistent state for exploring data without a cloud GPU.
Security & Honesty First
Guest-first design, rate limiting, secret scanning, no fabricated metrics — and every mock explicitly labelled.
Operational transparency
Libra runs on a strict zero-cost philosophy. Live inference uses real configured providers; anything simulated is clearly labelled. Status is always reported honestly — never fabricated.