Romain Zimmer
ML/AI Engineer, Paris, France (open to relocation)
Summary
ML/AI engineer with 7+ years building ML and AI systems and production-scale ML pipelines. Deep interest in AI algorithms and systems, model architectures, training methods, and infrastructure, alongside hands-on experience taking ML from research to deployment. Currently cofounder/CTO of a product observability platform for AI agents.
Experience
- Cofounded and lead technology for correl8, a product observability platform that turns AI agent conversations into structured product signals (observations, sentiment, grouped issues).
- Designed agentic processing pipelines with vLLM inference: an LLM-based ingestion pipeline (span classification, async issue digest, tag matching) and the Correl8 Assistant, an in-app conversational agent, plus background agents for autonomous analysis.
- Built MCP and REST integrations enabling third-party agents (LangChain, Pydantic AI, OpenAI Agents SDK) to post observations and conduct in-conversation research studies.
- Own full technical architecture and product direction as CTO.
- Hired to design and build a discount optimization system from scratch, enabling retailers to reduce waste and boost profits.
- Built and maintained data pipelines using DBT and Snowflake, and ML workflows using Metaflow.
- Developed predictive models for discount elasticity and a decision engine for optimal discounting.
- Built an internal experimentation framework from scratch for A/B tests of ML models; ran in-store tests with proven significant impact across multiple retailers.
- Delivered scalable ML infrastructure and CI/CD solutions.
- Refactored a Python codebase, implemented integration tests, and migrated a system from GCP to AWS Serverless.
- Deployed Triton Inference Server on AWS SageMaker and built Slack bots for ML validation tasks.
- Received funding via Inria Startup Studio to build Sievable, a machine learning-powered collaborative search engine.
- Created a scalable, deep learning-based search engine platform with user contributions and blockchain-backed incentives (token on Solana).
- Designed and deployed a full MLOps pipeline handling data ingestion, model training, embedding updates, and deployment.
- Conducted R&D on generative models for speech synthesis.
- Developed a Python library for benchmarking ML pipelines on ARM processors.
- Developed Spiking Neural Networks (SNNs) for audio analysis and modeled them using RNNs trained with surrogate gradient methods in PyTorch.
Side Projects
ML research on Jetson Orin Nano
-
SLMs Looped decoder-only transformer trained from scratch on TinyStories. Transformer architecture with RoPE, RMSNorm, and SwiGLU, trained with Muon + AdamW and deep supervision across inner loop iterations. Docker setup for training, text generation, and metric/sample visualization on Jetson Orin Nano.
-
Sudoku solver Looped MLP-Mixer sudoku solver with an outer commit loop, inner mixer iterations, and a learned halt head, trained from scratch on sapientinc/sudoku-extreme. Each outer step applies the previous prediction, refines the grid through inner mixer passes, and updates cell memory until the halt head signals a complete solution. Docker setup for training, eval, and trajectory visualization.
Consumer AI agents
- Teannamon Tea brand with a personalized AI assistant for recommendations and, soon, personalized tea creation.