Dynamic AI

Fine-tuned language models that run locally

Since August 2025 I have been the UX developer on Dynamic AI at Raw Power Labs, a platform for building small, fine-tuned, quantized language models that run locally on any device. We call the result a Diamond, and because it runs on the device itself it stays fast, private, sustainable, and free of any API dependency.

The models are used in two quite different worlds. In games they drive dynamic experiences such as character dialogue that reacts to what a player actually does, and in industry they power things like chatbots and email response generation, where sending data off to a hosted model is often not an option.

When I joined, the platform was a proof of concept built by data scientists, for data scientists. Everything it did was real, but the language, the structure, and the flow all assumed you already understood dataset generation, fine-tuning, and quantization. My job was to translate that into something a game developer, a writer, or a customer service agent could pick up.

That meant establishing the architecture, terminology, and visual hierarchy first, so the concepts had names people could hold on to. Everything else was built outward from there: a list-based template flow, a one-click Autoflow that produces a Diamond in a single step, and later a wizard that teaches each concept while you create your first project.

The most recent addition is graph-based dataset generation, where variables become nodes you can wire together and randomise on every run. It turned the part that was hardest to explain, how varied training data actually gets produced, into something you can see and change directly.

Raw Power Labs is small enough that the role has never been only design or only front-end. Alongside the platform I handled the visual identity and graphic work: logo design, marketing and promotional material, and the videography and editing that went with it.

The engineering side ran just as wide. I integrated authentication and identity, wrote the Cypress and Vitest suites and ran test management, and because the platform was aimed partly at games the work extended into game design and development alongside the engine integration itself.

Year

2025 · 2026

Role

UX Developer

Made for

Raw Power Labs

A company building tooling for small, fine-tuned language models that run locally on any device, without hosted APIs

Tools and stack

React, Next.js, Figma, Cypress, Vitest, Davinci Resolve, Godot, Unity, Unreal

Understanding the proof of concept

The existing platform was studied end to end, from template creation and dataset generation through to training and fine-tuning, so that every technical step it performed was understood before any of it was redesigned.

Architecture, terminology, and hierarchy

A project architecture and a shared vocabulary were established, and visual hierarchies were built around them, so that the concepts behind the platform could form clear mental models.

The first complete version

A complete version of the platform was built around a list-based template creation flow, giving structure to the prompting that the resulting models would later be trained on.

One-click Autoflow

An Autoflow was introduced so that a Diamond, a fine-tuned and quantized small language model, could be generated in a single step rather than assembled by hand.

Game engine integration

UX guidance was given on integrating the models into game engines, and support was provided during implementation so that the platform's output worked in the environments it was built for.

A wizard for new users

A guided project creation wizard was added, introducing each concept at the moment it was needed, so that newcomers could build their first Diamond without prior knowledge.

Graph-based dataset generation

Dataset generation was rebuilt as a node graph, making variable-driven generation visible and letting values be randomised on every run to produce varied training data.