Staff AI Architect — Agentic Shopping Lead

DeepCamp; Amsterdam; Hybrid; EUR 140,000 - 230,000; Full time

Benefits offered by this job

Relocation

25 holidays

Health insurance

Pension plan

Travel expenses covered

Phone and laptop provided

Job summary

Picnic is building the next-gen agentic shopping assistant to help households plan and shop effortlessly. You will own technical direction, architecture, and model strategy for multi-turn conversations, balancing latency, safety, and cost while mentoring engineers across the organization.

You will prototype, productionize, and iterate on retrieval, ranking, and context management using open-weight models, with a strong emphasis on evaluation loops and safety guardrails.

Qualifications

• Experienced in taking complex multi-turn conversational AI systems from design to production.
• Proficient in training, fine-tuning, and optimizing open-weight LLMs for deployment.
• Hands-on experience building production-grade RAG systems and context management.

Responsibilities

• Define and drive technical direction for agentic shopping experiences.
• Design robust architectures balancing latency, safety, and cost.
• Mentor engineers and lead design reviews across teams.
• Prototype, validate, and productionize new components and patterns.
• Collaborate with product and design to ensure safe, reliable user experiences.

Skills

Python

Open-weight LLMs

Multi-turn conversations

RAG

Production code

Model evaluation

Mentoring engineers

System architecture

Cost / latency optimization

Education

Degree in Computer Science / AI or related field

Tools

PyTorch

LangGraph

vLLM

Pydantic AI

Qdrant

DeepEval

Job description

Picnic is building the next-gen agentic shopping assistant to help households plan and shop effortlessly. You will own technical direction, architecture, and model strategy for multi-turn conversations, balancing latency, safety, and cost while mentoring engineers across the organization.

You will prototype, productionize, and iterate on retrieval, ranking, and context management using open-weight models, with a strong emphasis on evaluation loops and safety guardrails.