Global Texture - Noise Background
Global Texture - Noise Background
    Ecommerce banner site mockup
    Ecommerce banner site mockup

Project Details

Engineering reliable AI experiences

13 slide types

rendered deterministically — the model never touches the deck

0 ungrounded

numbers: every figure must appear verbatim in the source

Works offline

every model step has a deterministic fallback

Two systems

the product’s and the deck’s, kept deliberately apart

uxproof turns uploaded UX research into client-ready PowerPoint decks. Ask it in plain language — “Analyse Q3 2025”, “Compare Q2 vs Q3”, “Generate the 2025 presentation” — and it answers from your own uploaded research only, then renders the deck on demand.

A self-initiated concept, designed and built end to end: the interaction model, the design system, the React and TypeScript implementation, a multi-agent generation pipeline, and the slide template the decks are drawn from.

The first design artefact was not a screen — it was a boundary. The model may select which findings matter and narrate them. It may never touch the numbers, the slide structure, the styling, or whether the product works at all. Between the two sits a validation gate: every figure must appear verbatim in the source or it is dropped, and output that fails the check is discarded rather than repaired. Every model step has a deterministic fallback, so with the model switched off entirely, decks still generate.

Two design systems, deliberately separate. The application’s governs behaviour as much as appearance: where a control is allowed to live, what every state has to look like, how much emphasis a single view may spend. The deck’s is stricter and narrower — fixed slide masters, one accent colour with a job to do, flat shapes, no decoration. They share two typefaces and nothing else.

A working concept rather than a commercial product — real uploads, real decks, running on my own research practice, built to demonstrate an architecture.

Produced by

PAISAK4U

Role

UX engineering lead for the concept — interaction design, information architecture, design system, accessibility, React implementation, AI workflow and presentation renderer

Scope

Concept build, designed and engineered end to end: interaction model, design system, React implementation and a multi-agent deck pipeline

Year

2026

Tools

Next.js
React 19
TypeScript
Zustand
FastAPI
Pydantic
Sanity
Ollama
pptxgenjs
Figma
Claude Code
Design system — one filled button per view: variants, states and the rules that govern them
Design system — two families, eight sizes, and a rule against adding a ninth
Design system — the three non-content states, and a destructive flow that names what else goes
Design system — the assistant surface and the four states of the presentation card
Design system — every colour pair computed from the shipped tokens, and the rules that follow
Design system — the six surfaces a control is allowed to live on, and what each may hold
The same deck card at tablet width
The research behind a period, kept where it can be traced
File becomes evidence, evidence becomes a filed period — and every figure keeps the page it came from
The agent sequence, named as it runs — context, extraction, planning, render
A finished deck, ready to download
System architecture — explicit ownership boundaries, everything on one machine
One contract, three languages — the schema, the app and the pipeline share a shape
Deck spec — five colours, four flat ink tints, and the job each one is allowed to do
Deck spec — thirteen roles, each with its size, weight, tracking and leading
The generated deck, back in the tool that made it