AI Software Engineer at Replika · staff-level scope · LLM inference, agents & memory · Remote EU
France · remote across the EU
At Replika, I work on LLM serving for a consumer app used by millions. With vLLM and SGLang, I work through latency, GPU cost, and reliability alongside model quality and capacity. I also build agent infrastructure with tools, background tasks, and coordination across multiple agents, and contribute to a long-term memory layer.
Before this, I worked in data platforms and production machine learning. I designed and deployed custom embedding models for matching industrial parts. I have led teams of up to eight engineers, including at Stratio, where we brought time series models into production for fleet monitoring. Earlier still, my PhD research in Marseille explored numerical approaches to quantum gravity. Along the way I wrote open source C libraries, including sl2cfoam, with CUDA kernels for GPU parallelization.
Experience
Expand for contextLLM serving for a consumer app used by millions, weighing latency, GPU cost, reliability and model quality with vLLM, SGLang and third party APIs; agent infrastructure and a long-term memory layer.
October 2025 to present. Work on LLM inference, agent infrastructure covering tool use, background tasks and multi agent orchestration, and a memory layer.
Key initiatives
- Technical focus: inference, agents, and memory systems.
Designed and deployed custom embedding models for industrial part matching, and set direction for the data platform.
February 2025 to October 2025. Work on custom embedding models for industrial part matching and direction for the data platform.
Key initiatives
- Technical focus: custom embeddings and matching.
- Leadership: a team of four engineers.
Built an event driven recommendation system and the ML stack, including cold start matching and behavioral tracking.
Part time, 2024 to 2025. Work on an event driven recommendation system and the ML stack behind it.
Key initiatives
- Technical focus: cold-start matching and behavioral signals.
Led a team deploying time series models to production for fleet monitoring, and worked on automotive diagnostics.
2022 to February 2025. Work on production time series models for fleet monitoring and on automotive diagnostics.
Built route optimization tools for the marine industry and was responsible for product and technical direction.
2020 to 2023. Work on route optimization tools for the marine industry, with responsibility for product and technical direction.
Research and a PhD in theoretical physics on numerical quantum gravity, including open source C libraries such as sl2cfoam.
2017 to 2020. Developed open source C libraries, including sl2cfoam, using CUDA kernels for GPU parallelization.
Areas of focus
LLM inference
vLLM, SGLang, GPU sizing and routing to third party APIs.
Agent infrastructure
Tool use, background tasks and multi agent orchestration.
Memory systems and embeddings
Memory layer work and custom embedding models.
Technical skills
LLM serving
Agents and retrieval
Machine learning
Engineering
Writing
Notes on LLM inference, agent memory, and AI coding agents.
- AI coding agents make building cheap. Taste still matters.
- Agent memory after RAG: search, revise, forget
- Serving LLMs is a product decision