// the find
olivia-ai/olivia
💁♀️Your new best friend powered by an artificial neural network
Olivia is a Go-based chatbot that trains its own tiny neural network from a JSON intents file, classifies user input, and routes it to hardcoded modules (jokes, Spotify, math, reminders). It predates the LLM era and solves the problem LLMs now solve better by default. Worth looking at if you want to understand how intent-classification chatbots worked before transformers took over.
The neural network is implemented from scratch in Go with no ML framework dependencies — matrix math, forward pass, backprop, all hand-rolled, which makes it a readable learning resource. Multi-language support is genuinely structured: each locale gets its own intents, stopwords, and training data under res/locales/. The module system is clean — adding a new capability means dropping a file in modules/ and registering it, no framework ceremony. Test coverage is decent and CI actually checks formatting, not just build success.
The underlying model is a bag-of-words MLP trained on a handful of example phrases per intent — it will misclassify anything slightly off-script, and there is no path to improving it without rewriting the architecture. Last meaningful commit activity is from 2021–2022 despite the Feb 2025 push date; the call for contributors pinned in the README has been there for years with no resolution. The hardcoded modules (Spotify, movies, jokes) are shallow integrations that require API keys and break silently when those APIs change. Competing against GPT-4o or even a local Llama model for this use case is not a fight this architecture can win.