Senior AI Engineer - Freelance

About the job

Design, build, and scale production-grade AI systems that power the next generation of intelligent commerce. You will work at the intersection of Software Engineering, Machine Learning, and Generative AI to build AI-native products, agentic systems, and platform capabilities used by Mirakl customers and internal teams.

This is a freelance position based in Paris or Bordeaux, with 1 day of remote work per week.

Contexte : Mirakl, fondée en 2012 (French Tech Next40), édite un système d'exploitation de marketplaces (Mirakl Platform, Mirakl Connect, Mirakl Ads, Mirakl Nexus). Dual headquarters à Boston et Paris, plus de 750 salariés dans 9 bureaux, dont plus de 350 personnes dans les équipes tech principalement en France, 5 solutions SaaS, plus de 450 marketplaces clientes et un réseau de plus de 100 000 vendeurs tiers (Macy's, Decathlon, Carrefour, Asos, Airbus Helicopters). L'IA y devient un axe central du produit : Catalog Transformer, génération de contenu, agents IA et commerce agentique. Le poste porte sur les façons de travailler des équipes internes et sur le versant agentique des produits SaaS.

What you'll do

1. Design & Build AI Systems

• Design, build, and deploy production-grade AI systems for Mirakl internal operations.
• Develop solutions leveraging: LLMs, RAG, AI Agents, Multimodal AI and Evaluation frameworks.
• Architect complex agentic workflows including tools, memory, retrieval, planning, and guardrails.
• Rapidly prototype new capabilities and evolve them into scalable, reliable AI production systems.

2. Integrate Across the Stack

• Connect AI systems to internal platforms, business applications, and data sources across the company.
• Enable AI agents to interact with systems such as CRM, data platform, Drive, Jira, support tools, and internal knowledge bases.
• Design robust orchestration layers between models, APIs, tools, retrieval systems, and business workflows.
• Ensure reliability, observability, scalability, and maintainability across the entire AI stack.

3. Iterate for Impact

• Partner with end users, business teams, Product Managers, and Engineers to identify high-value opportunities for AI.
• Run pilots and experiments, gather qualitative and quantitative feedback, and measure business impact.
• Analyze usage patterns, failure modes, costs, latency, and performance metrics.
• Continuously improve accuracy, reliability, adoption, and user experience through data-driven iteration.

4. Ensure Quality, Security & Reliability

• Implement evaluation frameworks, guardrails, validation layers, and monitoring systems.
• Design AI systems that are secure, explainable, and production-ready.
• Ensure compliance with data security, governance, and access control requirements.
• Document architectures, capabilities, limitations, KPIs, and operational best practices.

5. Drive Adoption & Enablement

• Build a scalable ecosystem of AI systems and agents that empower both internal teams and customers.
• Support teams in adopting AI-powered workflows and integrating AI into their daily operations.
• Contribute to AI engineering standards, reusable frameworks, and best practices across the company.
• Act as a champion for AI-driven ways of working and help shape Mirakl's transition toward Agentic Commerce.

(La section « Our stack and tools » de l'annonce a été coupée à la collecte : la liste exacte des outils et technologies n'est pas reprise ici.)

Ce qui te ferait sortir du lot (côté sécurité)

Zeroday : sur un poste comme celui-ci, ce qui fait la différence côté sécurité :

• traiter tout contenu récupéré (pages, tickets, documents) comme non fiable face à l'injection de prompt, et valider les actions d'un agent avant exécution ;
• donner à chaque agent des jetons dédiés, à portée minimale, révocables et journalisés, jamais un compte de service tout-puissant ;
• masquer les données sensibles avant qu'elles n'entrent dans les prompts, et tenir les journaux d'appels exempts de contenus confidentiels ;
• tester les garde-fous comme le reste du système : non-régression sur les tentatives d'injection et sur les sorties interdites.