Uni matches what you say against rules it knows. Plain code answers the same way every time. A language model is asked only when the rules run out.
Watch the series on YouTubeGeorge Butiri built the first Uni as a voice companion to help manage ADHD. It ran on a Raspberry Pi for the first time on September 14, 2023, after a month of work.
Every conversation folds into a daily summary, the days fold into weeks, then months, then years. Uni keeps the shape of a long history without carrying all of it. That memory engine is Tapestry, published on PyPI as python-tapestry.
In April 2025 Uni changed course: rules first, model last. A matched rule runs plain code - fast, free and the same every time. Within a week Uni wrote a new intent, its pattern and its handler by itself: its first full self-programming cycle.
Parsely (on PyPI as parsely-dip) is Uni's front gate. A sentence tries a regular-expression rule first (microseconds), then a Stanza grammar parse (about 50 ms). Only if both miss does it fall back to a language model.
The Uni IDE runs on tuned local models - 14B, 7B and 3B - with Parsely in front. Work happens card by card: each task is committed, checked by a deterministic QA runner and merged when the card closes. The model fills in narrow steps; code decides what happens next.
The L.E.A.R.N. engine (Logically Evolving Adaptive Response Network) is designed and coming soon. Uni will grow one rule at a time, with a person approving each one until it earns more trust.
People author the patterns Uni matches.
A model proposes a new pattern; a person approves it.
Uni notices gaps and proposes its own patterns.
Now a person checks every new workflow. Next, automated QA (already running) takes on more. Later, Uni validates itself per task.
U.S. Patent Application 19/719,283: Deterministic-First Natural-Language Instruction Engine with Classifier-Free Library Routing and Runtime User-Confirmed Self-Expansion. Filed June 2026, claiming priority to July 2025.
Watch the series, follow along, or get in touch.