Nyx AI

What Nyx AI is

A conversational AI engine designed around an unfashionable principle: define the behaviour explicitly before you train anything. Where most modern AI is trained on enormous corpora and then patched into safety with RLHF and prompt engineering, Nyx starts with rules — explicit, auditable, edit-by-text-file — and treats machine learning as an extension to that scaffold, not a replacement for it.

Architecture

A C++20 rule engine with carefully-bounded scope:

• Normalised token and phrase matching — case-insensitive, configurable stemming, robust to typos and Unicode variants.
• JSON-driven intents — text patterns paired with response templates; intents are version-controlled text files, not opaque model weights.
• Lightweight conversation state — turn-by-turn, with explicit slot tracking.
• Persistent memory — a small per-user JSON store for facts the engine should remember across sessions.
• Scenario-based evals — every behavioural change is regression-tested against a scenario suite.
• Local usage metering — for users who want to know how much they’re using their assistant.
• CLI and desktop GUI entrypoints — a JSON-line server protocol lets any GUI talk to the same engine.

What you can do with it today

Nyx ships as a working CLI (nyx_cli.exe) and as a JSON-line server for GUI clients. The desktop GUI is in active development (Qt-style and WPF variants). The engine is suitable for constrained conversational tasks — research assistance, structured note-taking, task and reminder organisation — without invoking any hosted LLM.

Why this matters

This product is the working prototype behind our Rule-Before-Train AI Safety research — read that for the deeper argument about why the industry’s current approach to AI safety is structurally limited.

Maturity

Prototype. The engine is working and the eval suite is meaningful. The “delightful daily-driver assistant” story is still ahead — that’s the next year of work.