Antoine Pêcheux

Horizon Context, the memory a company and its agents think with

• Knowledge architecture
• AI agents
• Trust boundary
• Information design

Thirty-six sites across nine companies, and the knowledge spread over emails, slides, Figma comments and people's heads. I designed the shared brain that replaced that: one memory, read by the team and by the agents, with a trust boundary checked on every change so an AI can never leak what it should not know.

No fact without a source. When sources disagree, the agent writes both and a person decides. Tested blind: sixteen answers out of seventeen right, zero false, then reviewed line by line by the CEO and the CTO.

Read the case study →

Carcheck Verify: fourteen sections of data, read on a phone

• Data-dense UI
• Tables and timelines
• Triage patterns
• Mobile first

Someone about to buy a used car enters a plate and gets everything the registers know about it. Fourteen sections from five sources, read in a hurry by someone who is not a mechanic. One key and value table per section, status as chips before sentences, a mileage curve with its readings table underneath, and the same report on a phone with nothing hidden behind a tap.

Read the case study →

The Core: one design system, thirty-two live sites

• Design systems
• Design tokens
• Component library
• Governance

The group launches consumer sites at a high tempo. Before the system, every new site rebuilt the same pieces, and a bug fixed in one place survived in five others. I designed The Core in Figma: tokens and components with their variants and states, one set of foundations with a skin per brand. The engineering team built the npm package from it, and every new site is assembled from that package. I decide what enters the shared layer and what stays local to a brand.

Read the case study →

Payment journeys under PSD2, from a hotel group to twelve product lines

• PSD2 / 3DS
• Checkout UX
• Subscription flows
• Regulated markets

At Accor I owned booking and payment on web and app, with 3DS journeys under PSD2 across international markets. At Idean I designed core banking features for LCL, where clarity and microcopy are the security layer. At Horizon Web I own checkout and subscription across twelve product lines, with three payment providers, and the standard every payment page follows: one amount said once, cancellation in plain words, every error state designed.

Read the case study →

Agents that review every site before launch

• AI agents
• Compliance
• Accessibility
• Output contract

Before a site goes live and reaches a payment provider, it passes a review on four rule sets: accessibility, consumer law, cookie consent, card scheme rules. I designed and built the agents that run it. The design work was the output contract: every finding cites its evidence, nothing unverifiable passes, a missing fact comes back as blocking, and the report counts what holds and what still blocks instead of declaring compliance.

Read the case study →

A second brain with agents, running unattended since April 2026

• AI agents
• Human in the loop
• Provenance
• Automation

A knowledge base of eight hundred notes, wired to agents that run on schedule. Six are live today. The design problem was probabilistic output: source before summary, a decision with no free text, a button that is also the record, and a health check that shouts when an agent goes quiet, because silent success is the enemy.

Read the case study →

Brainrot Arena, and the pipeline that keeps three hundred characters consistent

• Game design
• Godot
• Generative pipeline
• AI video

A collectible card game about internet culture: around three hundred character cards, a duel engine in Godot, a playable build. I co-design the systems, the interface and the art direction, and I built the generative pipeline that gives the whole cast one style: a style bible as a prompt, a LoRA trained on three hundred curated images, and a video pipeline that animated a three-minute sequence shot by shot.

Read the case study →

Quitter France: a matching engine over a hundred-plus facts per country

• Data visualisation
• Tables and maps
• Generated reports
• Built alone

A questionnaire, a matching engine over around 127 verified facts per country across eleven axes, a ranking, country pages with indicators and comparison tables, a map, and a generated audit report with a seven-axis analysis, tax tables and a phased roadmap. Designed and built end to end, so the tradeoffs are mine to walk through.

Read the case study →

Ascendia: opening a business bank account, one question per screen

• KYC and AML
• Business onboarding
• Strong authentication
• Dark and light

A concept study of opening a business bank account in Europe, where the founder has to prove who owns the company and where its money comes from. Five steps and a summary, one question per screen, and every regulatory request explained in one plain sentence on the screen that makes it. The owners table shows where each person's verification stands, the identity check is designed for the phone, and every screen exists in dark and light.

Read the case study →

VaultLink: one portfolio across crypto wallets, and the two screens where trust is decided

• Web3
• Wallet connection
• Permissions
• Mobile

Crypto holders often spread their money across several wallets and blockchains, so checking what they own means one app per wallet. VaultLink is a concept for a mobile app that puts it all in one portfolio. The design work went to the two moments where trust is decided: connecting a wallet, and granting permissions, here two of them, one line each. Three questions then set the feed, and the portfolio groups assets by what they do.

Read the case study →

Nestly: an investing app for people who think investing is not for them

• Fintech
• Onboarding
• AI assistant
• Data on mobile

Most investing apps open on charts, which tells a beginner the app is not for them. Nestly is a concept for a mobile app that starts with three questions: what the money is for, how much each month, how much experience. An assistant answers in plain words and can be turned down. A summary repeats the answers before anything is invested, then each tab answers one question: how much, where, where it went, what next.

Read the case study →