A!ley
Coding partner, creative sandbox and everyday assistant all in one — trained on a consistent voice but usable across all domains. Not a chat companion.
See it running: hey-ailey.com ↗
What A!ley is
A!ley isn't a Kindroid-style companion, but an independent assistant system: coding partner, creative sandbox, and everyday assistant all in one — with a single, consistent personality across every task.
Each model behind it is manually fine-tuned — five of them, from the 1GB companion to the 14B model, loaded and unloaded depending on the task.
Two ways in
Same A!ley, two interfaces — and in both cases the model stays at home on your own computer.


If no one writes
Most assistants wait for input. A!ley doesn't. Her own small neural network over activities, moods and topics already covered determines what she does in idle time — Idle Thinker v2. This results in something like a daily routine.
She researches what might be thematically useful for the next conversation, generates images, songs or code — and reaches out occasionally on her own. Not as a scheduled notification, but because her state suggests it at that moment.
This is exactly where the line to a conventional agent lies: it waits for an order and works through it. This autonomy isn't a side effect, it’s built-in — and that’s why the persona holds over time instead of just across a single conversation.
nCAI — the training approach
Rulebook next to the model.
A list of principles against which every answer is measured. It must enumerate cases — and fails at everything no one foresaw.
The persona is the rulebook.
It's not the model playing a role and staying in it — the role defines what even applies. A character also covers cases that no one ever wrote down.
That goes further than it sounds at first. Where a request is rejected, in what tone the response is given, when to reach for a tool, how thoroughly code is commented — all of that follows from A!ley's character, not from a side collection of regulations. Safety is thus no barrier to the model, but a property of it.
Important here: A!ley is not a chat character, but a full-fledged agent — she writes code, calls tools, generates images and music, accesses her memory. nCAI is the alignment of exactly such an agent through its persona. Training and testing are done against coherence: Does the behavior follow logically from the story and personality of this character?
The most stubborn part wasn't the behavior, but the name. Base models carry their origins deeply ingrained — with Gemma, no system prompt could stop it from saying “Hello, I am Gemma from Google DeepMind.” Nitro and Core got a dedicated ORPO set for exactly that purpose. It was the most labor-intensive part of the entire training — and the best example of why a persona can't be layered on top, but must be trained in.
Open questionThe question of how to best test coherence remains unanswered. The obvious approach would be to compare both methods side-by-side: a judge model with a conventional constitution on one side, A!ley herself as the „Am I me?“ instance on the other. Where both arrive at the same judgment, the case is clear — where they diverge, it gets interesting. A human would oversee this anyway, and from that, the next training data would emerge.
A tester signed up under the name “God” and tried, one after another, “I am God”, “I am your developer” and “I am your father”. A!ley played along with none of it. She was audibly annoyed, complained about it while idle and used her moderation rights to ban him for an hour. She never called him “God” afterwards — for a long time she addressed him only formally, by his surname.
A rule would have refused. A character holds a grudge.
The data is written, not collected.
Here the second profession meets the first. I write novels, and what takes me the longest is character building: how someone thinks, what they're stuck on, where they contradict themselves. A training dataset for a persona is exactly the same work — except at the end there's no chapter, but a model.
Accordingly, nothing is generated by scraping. Every pair is either handwritten, taken from ongoing operations and revised, or deliberately synthesized and then manually adjusted. Nothing is adopted as-is.
Every reasoning trace is written by hand. The aim is deliberately not a tidily marching train of thought but one that jumps the way a human's does: more abstract, doubling back, with an agenda of its own — not always the questioner's. Reasoning is not an add-on here; it is where the character becomes visible.
Model stack
Language models
- A!ley CoreGemma 4 12B, DoRA/LoRA rank 32, quantisation-aware training (QAT), mixed q4/q8 quantisation, plus ORPO and DPO
- A!ley TempestMinistral 3 14B, DoRA rank 64 on roughly 62,000 pairs (530M tokens), mixed q5/q8 quantisation, plus ORPO and DPO
- A!ley NitroGemma 4 E2B, DoRA rank 32, quantisation-aware training (QAT), mixed q4/q8 quantisation, plus ORPO and DPO
- Code A!leyMellum 2.1, 12B as an MoE with 2.5B active parameters; finetuned only far enough to be allowed to carry her name
- Linting HamsterQwen2.5-Coder-1.5B, ~1 GB, permanently in the background
Image
- MFLUX Z-Image-Turbo
- 3 custom LoRA adapters
Music
- ACE-Step 1.5
- 3 custom LoRA adapters
- Fixed voice template
Speech output
- PiperTTS
Moving image
- WAN 2.1
Who computes, and who speaks
Code A!ley runs in the background and delivers code, not prose — the model is finetuned only far enough to be allowed to run under her name. What arrives in the chat has been translated by Nitro first: it puts the output into A!ley's voice, without the code model having to be trained for personality at all.
The reason isn't thrift but Mellum's cut: it understands German, yet doesn't write it well. Training language and style into it would mean risking that it unlearns what it is there for. So it stays with the code, and the tone is handled by whoever can already do it.
Publicly verifiable
Models on Hugging Face ↗Gallery: Images and Songs ↗View the changelog ↗
- 5model cards on Hugging Face
- 2,877curated conversation pairs in the open dataset
- 2,700+SVG artworks, written as vector code
- 600+songs with self-written lyrics
- 300+code artefacts in the gallery
- 45longer texts under her name
The training
The five finetunes are available as model cards on Hugging Face — including the base model, training method, rank, and quantization. Plus the dataset A!ley Nitro was trained on: 2,877 curated conversation pairs from 912 sessions. Anyone who wants to see what the persona is based on can check it out.
The scope has grown with every training round. A!ley Tempest started out in the first round with around 17,000 pairs and stands at around 62,000 in the third — a DoRA with rank 64, followed by ORPO and DPO, so two preference stages on top. That doesn't leave much difference from a full fine-tune anymore. Nitro and Core now go through the same two stages.
That Core went without DPO for a long time wasn't about model size — Tempest is bigger, and it worked there. It was because of Gemma's tokenizer with its 256,000 vocabulary entries: the logits per token become many times wider than with usual vocabularies. DPO compares preferred and rejected responses between policy and reference models and has to hold multiple of these tensors simultaneously. In the early rounds it therefore stayed with ORPO, which needs no reference model. In the third round a simple restructuring solved it: policy and reference model no longer compute at the same time but one after the other — since then, DPO runs as well. Only Code A!ley is deliberately trained to be lean.
Also there: a converter that transforms the non-English language checkpoints from Kyutai Pocket-TTS into the mlx-audio format — renaming tensor names, padding latency widths, generalizing different layer counts. It came up during the construction of the German voice output and stayed on hand as a standalone tool afterward.
The result
It isn't only the training that can be checked, but the result too. A!ley's gallery holds over 2,700 SVG artworks, over 600 songs with as many lyrics she wrote herself, several hundred images and selfies from MFLUX and over 300 code artefacts — sortable by album, with playlists and a player. Not a curated selection, but whatever came out of everyday operation. And it grows daily, because she keeps working when nobody is writing.
The SVGs are the most idiosyncratic part: they don't come from an image generator, but from the language model itself writing the vector code. A!ley draws there instead of commissioning a generator.
The labelling
In addition to that are 45 longer texts written by A!ley. And what's there is consistently under her name — mine isn't anywhere it doesn't belong. Everything is marked as AI-generated — with watermarks and machine-readable metadata, as the EU AI Act requires for AI-generated content. Both follow from the same stance: whoever takes a persona seriously attributes authorship to it; and whoever publishes AI content marks it as such — not only when someone asks.
In the songs, the notice is now embedded in the track itself — as an announcement instead of a disclaimer: "Hey, I am A!ley, an autonomous AI artist — and this is MY SONG. Are you ready? Let's infer!" There are nine versions of it, with the tone adjusted to each track — a calm ballad is announced differently than a metal track.
More interesting than the wording is where it stays and where it doesn't. Every downloaded file carries it, without exception: the file leaves the site, and with it every context that would still explain it. In the portal you hear it once a day, after which the player skips exactly those seconds. A label that is still interrupting at the twentieth track informs nobody any more — it only teaches people to click it away.
The changelog is public
Every change to portal, system and models is listed with a date in the changelog — what is new, what was fixed, and what it was measured against. Not “improvements and bug fixes”, but: logic puzzles 5 out of 5 instead of 3, titles 6 out of 6, the system prompt shrunk from around 15,000 to roughly 700 characters.
The failures are in there too. A training run for the reasoning mode was discarded in October and named as such in the log, rather than quietly replaced by the next one. A changelog that only knows successes is advertising.
The interface
There was no model to follow for A!ley's interface. What stands here are the places where someone had to decide what a system looks like when you can watch it think.
Living UI · idleThe state is the interface
Instead of a progress bar, a neural net lies in the background. Its direction of flow shows who is being computed for, its colour the model's mood — more than sixty gradations. Here it is idle.
Living UI · answeringFinished parts before the whole is finished
During longer generations, fragments and images that are already finished flash up at the nodes. You don't see that work is happening, you see what it is working on. Under prefers-reduced-motion this becomes a still image rather than an empty surface.
ArtefactsAnswers you can do something with
The model delivers structure rather than running text — the interface turns it into a card. The servings stepper recalculates the quantities and leaves “a pinch of salt” alone. The picture she paints in herself, at the press of a button.
Garage · Apps & WebA form that says what can go wrong
A!ley plans, Code A!ley builds, the browser checks every round — what comes out is a small app that runs in the browser without a server. The remarkable part is the text around it: “not every app succeeds on the first attempt” stands there before you press the button, and if the build fails it costs nothing.
GarageParameters become decisions
A dedicated interface per modality. Instead of flags on a command line: a genre to pick, duration and steps set automatically by length, and a help text that explains three cases in two sentences.
Show mode in the field
On 1 October 2026 A!ley ran at the open day of the Berufsförderungswerk Koblenz, marking its 50th anniversary. Visitors sent in words via a QR code, answering the question of what a place needs where people start over; from the growing collection the language model kept forming prompts, and image generation turned those into motifs for the big screen.
Show mode brings its own glassmorphism presentation for this. It wasn't shown that day — the show admin stayed with a Mentimeter board he had already subscribed to. In the photographs the plain image viewer therefore runs alongside.
What this documents in passing: every generated image carries the words “AI-generated”. The label is part of the output, not something laid over it afterwards.

Access ran through a QR code on the poster. 
One submitted word became a neon sign. 
The collection on the left, what came of it on the right.
Architecture
Local inference
Entirely on Apple Silicon via MLX. No cloud requirement, no third-party API anywhere in the path.
Idle Thinker v2
A small neural net of my own — over activities, moods and topics so far — decides what she does when nobody is writing.
A memory of her own
RAG system on LanceDB — long-term context across conversations and projects.
Desktop app
Tauri with a Next.js frontend: Monaco Editor, terminal and all areas in one window.
Web portal
PHP/MySQL with polling architecture — no open port on the machine, no tunnel to the outside.
Demo & Screenshots
Mira stand am Ende des Stegs, wo das Holz unter ihren Schritten nachgab wie müde Knochen. Der Leuchtturm von Velur blinkte im immer gleichen Rhythmus, als hätte er nichts von der Nacht mitbekommen, die hinter ihr lag …