Departures

Aurélien Duveau, AI Engineer

Departed
2012
Gate
Status
Open to offers

AI EngineerLLM agents, tools and evaluation, in productionSeine-et-Marne, east of Paris (FR)

I design AI agents and tools, and I ship them to production. At Nautil: a multi-agent assistant (Claude) that diagnoses production bugs, and TarifScope, which checks prices against suppliers and 11 competitors. Behind that, over ten years in tech: dev, support, QA, then project management.

  • AI agents and tools in production
  • Agent evaluation with StayBench
  • Dev, QA and project management, in English

AI at work

Tools built with AI

AI is at the heart of how I work. With Claude Code, I take a business need all the way to a tool in production: I frame the need, design, have the code written, review it, test it and deploy it. Two examples below, with demos replayed on fictional data.

Nautil Helper

Built with Claude (Anthropic SDK), Python, FastAPI, React, TypeScript

The need
Understanding a production bug in the hotel connectivity stack meant cross-checking several code repositories, each supplier’s documentation and the database. Slow, and only possible for people who knew the whole ecosystem.
What the tool does
Nautil Helper is a multi-agent AI assistant I designed and built on my own initiative. It answers the team’s technical questions, diagnoses production bugs and writes the tickets.
specialised agents, coordinated by an orchestrator
4
code repositories covered (.NET, React, back office, docs)
6
suppliers documented, PDFs converted to Markdown
4
Claude models, picked per task
3
Multi-agent diagnosis Fictional data

helper diagnose booking-api.log

[ERROR] 09:42:11 BookingService Supplier B: price mismatch on booking 48213 (contract 1240 EUR, supplier 1410 EUR)

Orchestrator question routed to Code Expert

Code Expert reads the Supplier B connector passenger mapping and their API docs

Diagnosis non-standard OTA child age code, counted as an adult: the supplier price is wrong

Ticket Agent ticket written: cause, files involved, proposed fix, test case

Diagnosis complete

Fictional log, booking and documentation, to show how it works.

In detail

  • An orchestrator routes each question to the right agent.
  • A Code Expert agent explores the code: file reading, ripgrep search, supplier API documentation.
  • A QA agent reviews code changes.
  • A Ticket agent writes structured tickets.
  • Cross-analysis of code, supplier docs and SQL Server queries. Real cases: price gaps between connectivity and contract, non-standard OTA age codes, passenger types overwritten on the front end.

How I built it with AI

  • Multi-agent architecture with the Anthropic SDK and Claude tool use.
  • The right model for each agent (Opus, Sonnet or Haiku), balancing cost and quality.
  • Knowledge base: Beachcomber, Koedia, Tahiti Nui Travel / Tourplan and Juniper docs converted automatically from PDF to Markdown, plus a knowledge sheet written for each repository, injected into the agents’ context.
  • FastAPI back end (Python 3.12, async, SSE streaming), React, TypeScript, Vite and Tailwind front end, SQLite persistence.
  • Workflow industrialised with Claude Code: Jira ticket creation (REST API v3), Bitbucket access, Microsoft Teams notifications (Adaptive Cards).

TarifScope

Built with Claude Code, Python, Flask, Docker, Playwright

The need
Checking that hotel prices were accurate and competitive was done by hand: booking engine, suppliers and competitors, one by one.
What the tool does
TarifScope, an internal pricing web app, runs these checks automatically and alerts the teams as soon as a gap appears.
hotels in the connectivity guide
2,800
connectivities and bed banks covered
6 + 6
competitor sites tracked by destination
11
quotes and bookings replayed at current prices
Daily
Price check Fictional data
Demo table: hotels, suppliers and prices are fictional.
Offer Internal price Supplier price Gap
Hôtel Lagon ClairSupplier A 412 € 412 € …
Pension Motu ItiSupplier B 186 € 198 € …
Résidence CorailSupplier C 264 € 263 € …
Villa Sable NoirSupplier A 530 € 530 € …
Hôtel des AlizésSupplier D 158 € 158 € …
Lodge Vallée HauteSupplier B 342 € 309 € …
Hôtel Baie TurquoiseSupplier C 221 € 224 € …
Bungalow RécifSupplier D 475 € 475 € …

Click "Run the analysis" to compare prices. A simplified version of the supplier check.

In detail

  • Compares booking engine rates with supplier connectivities (Koedia, bed banks, local DMCs).
  • Also benchmarks prices against 11 competitor sites, destination by destination.
  • Replays quotes and bookings daily at current prices to catch drops and inconsistencies.
  • Sends automatic summaries by email or Teams to sales and product teams.
  • Checks hotel data quality (rooms, content) across reference systems.

How I built it with AI

  • Designed and built with Claude Code, from the teams’ need to production.
  • Python and Flask back end, REST APIs documented with Swagger, parallel calls to supplier services.
  • Playwright and Chromium to collect competitor rates and check content.
  • Scheduled jobs, email and Teams alerts, real-time progress (Server-Sent Events).
  • Docker, separate staging and production, Git with pull requests and release tags.

What I set up with AI

  • Business tools in production

    TarifScope, designed and built with Claude Code, checks prices against suppliers and 11 competitors, and alerts the teams by email or Teams.

  • Specialised AI agents

    Nautil Helper: 4 Claude agents (orchestrator, code, QA, tickets) that diagnose production bugs across 6 repositories.

  • Evaluating AI agents

    StayBench, my open source sandbox: 29 hotel booking scenarios with injected supplier faults, to measure whether an AI agent books correctly over MCP.

  • An industrialised workflow

    With Claude Code: Jira tickets created automatically, Bitbucket access, Teams notifications.

  • Working with coding agents

    Mypixelteam, my open source app, shows Claude Code, Codex and Gemini agents at work in pixel art.

  • A complete web product

    Vigie, built with Claude Code: monitors the health of a company’s customers and suppliers from French public data (BODACC, Sirene), live at vigie.soclyse.fr.

  • A game, from concept to release

    PatchThrough, built with Claude Code: an idle browser game in French and English, with fully tested game logic.

Itinerary

Journey

Seven stops, from work-study to today. Most recent first. Open a boarding pass for the details.

  1. NTL 06.2025 → now

    Current

    Boarding passNTL 2025

    Passenger
    Aurélien Duveau
    From
    June 2025
    To
    Present
    Company
    AAM Communication, on assignment at Nautil. Travel agency, hotel aggregation platform

    I joined as a QA engineer and soon took over as project manager on the hotel aggregation platform, which brings together offers from 4 suppliers: Juniper, Koedia, TahitiNuiTravel and Beachcomber. I work almost entirely in English.

    On board

    • I set priorities, run the dailies and follow every delivery through to release.
    • I represent the team with business stakeholders and suppliers: discussions on their APIs, technical trade-offs.
    • I test and sign off every delivery before deployment.
    • TarifScope: a web app (Python, Flask, Docker) I designed and built with Claude Code. It compares booking engine prices with suppliers and 11 competitor sites, replays quotes and bookings every day, and alerts sales and product teams by email or Teams. Sales staff and agents no longer check prices by hand.
    • Nautil Helper: a multi-agent AI assistant (Claude) I designed and built. It answers technical questions across 6 repositories, diagnoses production bugs by cross-checking code, supplier docs and SQL Server, and writes the tickets.
    • Set up CI with automated API tests (Jenkins, Postman/Newman) and reporting in Teams.
    • Data flow map documented in Confluence, .NET 8 endpoints, UI/UX redesign of the agent interfaces.
    • Preparing the migration from .NET Framework 4 to modern .NET, with a test strategy.
    See the tools built with AI

    Stack

    • Claude Code
    • Anthropic SDK
    • Python
    • FastAPI
    • Flask
    • React / TypeScript
    • Docker
    • Playwright
    • SQL Server
    • Jira
    • Bitbucket
    • Jenkins
    • Postman / Newman
    • Teams
    • Confluence
    • .NET 8
  2. PRJ 07.2024 → 06.2025

    Boarding passPRJ 2024

    Passenger
    Aurélien Duveau
    From
    July 2024
    To
    June 2025

    A year to focus on my own projects: development on FiveM, the multiplayer server platform for GTA V, and various other projects.

  3. CTS 10.2023 → 07.2024

    Internal move

    Boarding passCTS 2023

    Passenger
    Aurélien Duveau
    From
    October 2023
    To
    July 2024
    Company
    Cityscoot. Electric scooter sharing

    I managed the QA team and followed each person's progress, until Cityscoot closed down after going into liquidation.

    On board

    • I defined the test strategy.
    • I planned test campaigns around delivery dates.
    • I traced bugs to their root cause and proposed fixes.
  4. CTS 02.2023 → 10.2023

    Internal move

    Boarding passCTS 2023

    Passenger
    Aurélien Duveau
    From
    February 2023
    To
    October 2023
    Company
    Cityscoot. Electric scooter sharing

    I moved from development to quality, in the same company.

    On board

    • Automated functional and UI tests (Appium, Robot Framework, Selenium).
    • API testing with Postman, test cases in TestRail.
    • Coordination through Jira and Confluence.

    Stack

    • Appium
    • Robot Framework
    • Selenium
    • Postman
    • TestRail
    • Jira
    • Confluence
  5. CTS 01.2018 → 02.2023

    Boarding passCTS 2018

    Passenger
    Aurélien Duveau
    From
    January 2018
    To
    February 2023
    Company
    Cityscoot. Electric scooter sharing

    Five years between code and support, close to the users.

    On board

    • Developed the CRM (CakePHP 2) and the customer website (CakePHP 3), on MySQL and SQL Server.
    • Technical support on Freshdesk: analysis, fix or escalation through Jira.
    • Managed internal access rights.
    • Automation with Python and Airflow, documentation in Confluence.

    Stack

    • CakePHP 2
    • CakePHP 3
    • MySQL
    • SQL Server
    • Freshdesk
    • Jira
    • Python
    • Airflow
    • Confluence
  6. AAM 2015 → 2017

    Boarding passAAM 2015

    Passenger
    Aurélien Duveau
    From
    2015
    To
    2017
    Company
    AAM Communication. Web agency

    Client websites for a travel agency, a public body and a real estate agency.

    On board

    • Worked directly with clients, from gathering needs to delivery.
  7. CLO 2012 → 2015

    Boarding passCLO 2012

    Passenger
    Aurélien Duveau
    From
    2012
    To
    2015
    Company
    Clio Voyages Culturels. Cultural tours

    First steps in the travel industry, already.

    On board

    • Internal management intranet (PHP, ASP).
    • User support, Windows Server, hardware.

    Stack

    • PHP
    • ASP
    • Windows Server

Baggage

What I bring

Each point links back to a stop on the journey.

  1. Mentoring and growing people

    I've always done it. At Nautil and at Cityscoot, I trained work-study students and supervised permanent staff as well as contractors. At Cityscoot, I also managed the QA team.

    See the stop CTS
  2. Bridging tech and business

    Years of development, support and QA: I understand developer constraints and can explain them to business teams or suppliers.

    See the stop CTS
  3. Communicating in French and English

    English is my everyday working language at Nautil, with the team and with suppliers.

    See the stop NTL
  4. Building tools that matter, with AI

    TarifScope, an app in production built with Claude Code. Nautil Helper, a multi-agent assistant built on the Anthropic SDK. From need to tool, without waiting for a team to free up.

    See the stop NTL
  5. Shipping solid work

    My QA background makes me test and sign off every delivery before release.

    See the stop CTS

Tools

  • Claude Code
  • Anthropic SDK (agents, tool use)
  • Python
  • FastAPI
  • Flask
  • React / TypeScript
  • Docker
  • Playwright
  • SQL Server
  • Jira
  • Confluence
  • Postman
  • Jenkins
  • .NET / C#
  • PHP

Recommendation

What my former CTPO says

Aurélien grew brilliantly at Cityscoot, moving from apprentice developer to QA Lead, proof of his commitment and his ability to adapt and grow with the company. His path reflects his determination, his loyalty and his constant desire to learn, qualities that made him an essential pillar of our team.

Always ready to support his colleagues, Aurélien stands out for his generosity, his discipline and his attention to detail. His move to QA Lead was marked by deep expertise in testing mobile applications and APIs in a challenging context combining software and hardware, showing his versatility and his ability to take on complex challenges.

I recommend Aurélien without reservation for his professionalism, his team spirit and his ongoing commitment to learning. His rich experience as QA Lead, combined with his kind and collaborative nature, makes him a valuable asset for any organisation looking to strengthen its team with an exceptional technology and people leader.

Benjamin L. Former CTPO, Cityscoot LinkedIn recommendationTranslated from French

Education

Education and languages

  1. 2017 → 2018 Software Architect (Master’s level) CFA Insta
  2. 2015 → 2017 Software Analyst CFA Insta
  3. 2012 → 2015 BTS SIO (2-year IT degree) CFA Insta
  4. 2011 → 2012 Baccalauréat STG, Information Systems Management

Languages

French
Native
English
Fluent, working language

Off duty

Side projects

Complete products, built in my spare time.

  • Live

    Vigie

    Monitors the financial health of a company's clients and suppliers using French public data (BODACC, Sirene). Part of the Soclyse ecosystem, alongside Paraphe.

    Visit vigie.soclyse.fr
  • Live

    StayBench

    Open source sandbox that tests AI agents on hotel booking over MCP. It simulates a supplier, injects the failures of real distribution (price change at prebook, room sold out between two calls, a timeout hiding a booking) and grades every run deterministically, with no LLM judge. 29 scenarios, replay-verified leaderboard.

    Visit StayBench
  • In progress

    Mypixelteam

    Open source desktop app showing AI agents at work in a pixel art open space. Connected to Claude Code, Codex and Gemini.

    Visit MyPixelTeam
  • Live

    PatchThrough

    A small idle game in the browser: night operator at a 1950s telephone exchange. You patch calls through by hand, then apprentices and automatic switches take over, until the position is gone. French and English, mobile and keyboard, fully tested game logic.

    Visit PatchThrough

Final call

Boarding

Have a role to fill? Email is the simplest way.