Clément
Farkas Daudu

Front-end Technical Lead & Architect. I design the front-end architecture of SaaS products and make AI-assisted development work for a whole team.

Now
Stonal, since 2023
Team
10+ people
Focus
Design systems, micro-frontends, AI tooling
Status
Open to opportunities

Work

On a real-estate SaaS platform, for a fleet of independently deployed React apps.

01 — Design system

A React design system for the whole fleet

Designed and led for a dozen apps. A deliberately small API, composition over props, and the versioning discipline that keeps one component change from breaking every consumer at once.

30+ components built on the same compound pattern

  • Storybook
  • React Aria
  • TanStack
  • accessibility tests
tokens → componentstry the switch
surfacetextborder
02 — Architecture

One composed platform with Module Federation

A fleet of about 13 separately deployed React apps brought into a single shell: no reload between apps, shared dependencies as singletons, and every app keeps its own deployment.

12 apps normalized in five days against an estimated two to three months; 8 already live in the shell

  • Module Federation
  • codemods
  • independent deploys
one shell, many apps0 reloads
shellapp 1app 2app 3
App 1loaded on demand
react · singletonreact-dom · singletondesign system · singleton
03 — AI tooling

AI in the pipeline, with guardrails

Agents, skills and hooks, automated PR review, CI integration, multi-repo orchestration. Most of the work is in the guardrails that stop a confident wrong answer from reaching a branch, not in the prompts.

682 agent sessions measured to find where the cost really goes

  • Claude Code
  • agents
  • hooks
  • CI
every step is a guardrailrun #1
  1. agentproposes
  2. hookscheck
  3. PR reviewcomments
  4. CIvalidates
04 — Foundations & team

Shared libraries, fleet migrations, team

Front-end types derived from OpenAPI specs, auth, tracking, an app template. Migrations go through one canary repo in real CI first, then fan out to the fleet. Plus leading a team of 10+ people: conventions, upskilling, code review.

AI in production, not in a demo

Making AI-assisted development reliable at team scale: guardrails, measurements, and humans who stay in charge.

  1. PR review
    AI pull request review across the front-end fleet.Rolled out to 18 repos, with a stronger model on the 4 most critical libraries. Median cost: $0.51 per review. The review fails closed: no verdict, no green check. Its recall was audited against human reviewers over 1,832 PRs merged in 30 days.
  2. Learning
    A self-improving loop, human-gated.Review findings feed a ledger (2,104 records, 844 reviews, 16 repos) that proposes skill updates; a human approves every PR. 19 of 23 were merged.
  3. Skills
    The team's skill catalog.Co-maintainer of the internal Claude Code marketplace: front-end review, fleet migration and app scaffolding skills (about 80% of the boilerplate generated). Added evaluation and routing gates in CI for a 100+ skill catalog, 81 of which had no tests. Used by about 80 engineers over a month.
  4. Guardrails
    Guardrails over prompts.28 versioned hooks: the main agent orchestrates and only subagents edit code, every agent must declare its model, destructive git operations are blocked. Nine hooks merged into one dispatcher, proven by a differential test: up to 87 s saved per session.
  5. Orchestration
    Parallel agents, canary repo first.14 repos migrated in waves of 3 to 4 subagents. A design system audit by 34 reviewer agents, with a refutation pass, led to a 42% smaller bundle, 40% less JS in the host app and a 72% smaller published package.
  6. Memory
    Memory that survives sessions.An Obsidian knowledge base fed automatically by hooks (commits, decisions extracted from sessions): about 1,400 notes including 320 auto-captured decisions, injected at the start of every session.
  7. Measurement
    Measure before arguing.Latency studied over 20,194 agent turns: 4.9 s median under 100k tokens of context, 30 s above 800k. It tracks context size, not hook count. Token cost measured over 682 sessions.

Writing

Things that cost me a day, written down so they cost you less.

All notes on GitHub

Open source

Coding-agent skills, published with their before/after evaluation.

github.com/cdfarkas

cdfarkas-skills

Skills whose behaviour is a script, not prose to interpret. Each one is run against an agent without it, on the same prompts, graded by regex, and the results are published as they happened, failures included.

Experience

From web design in 2002 to technical lead.

  1. 2023 – now
    Stonal — Front-end Technical LeadFront-end architecture of a real-estate SaaS platform, leading a team of 10+ people.
  2. 2020 – 2023
    La Place de l'Immobilier — Front-end Lead, then Technical LeadFront-end quality and architecture, Storybook design system, training; then onshore and offshore management, back- and front-end design, devops.
  3. 2019 – 2020
    Mooncard — Front-end LeadNew component-based front-end architecture, quality standards and tooling.
  4. 2019
    Deezer — Front-end Lead, Core ApplicationsA team of five senior engineers on the core libraries of the web apps, front-end community of practice.
  5. 2017 – 2018
    Keyrus — Front-end LeadWeb performance and front-end optimisation for e-commerce sites, coordinating offshore developers.
  6. 2016 – 2017
    Abbeal — Senior front-end consultantIndustrialisation, code quality, web performance.
  7. 2008 – 2016
    Cellfish — from integration to managementMultimedia integration, mobile web project lead, then manager of the front-end developers.
  8. 2002 – 2008
    Orange R&D, web agencies — multimedia expert, web designer

Elsewhere

Off screen, and education.

  1. first aid
    Volunteer first aider with the French Red Cross (PSE2)
  2. stage
    Amateur theatre with the Les Improvisibles troupe
  3. 2003
    Web design, ITECOM
  4. 2002
    Two-year degree in business computing, developer track

Let's talk.

l4ir.fr · updated October 2026 · no trackers
l4ir.frFR

Front-end Technical Lead & Architect · Paris area

Clément Farkas Daudu

Design systems, micro-frontends, and AI-assisted development that holds up in production. Currently at Stonal.