Research Scientist, Agentic AI for Closed-Loop Scientific Discovery (Founding Team)

Place: Paris, France. Level: Entry level. Contract: Permanent. Mode: Hybrid.

Snapshot

We're looking for a research scientist to help build agentic systems for closed-loop scientific discovery with autonomous agents that hypothesize, plan, ground, and act across long horizons to discover new thermal materials.

About Us

Thermia is building multi-scale foundation models to discover the next generation of PFAS-free thermal fluids and systems. Our AI team is a tight-knit group of ML scientists, computational chemists, and thermal engineers building generative, RL-based, and agentic systems for molecular and physical discovery. We are an early-stage company based in Paris, working at the intersection of agentic AI, scientific ML, and industrial thermal design.

The Role

As a Research Scientist on this team, you will design and build the autonomous scientific agents that drive Thermia's closed-loop discovery. You will engineer how they read the literature, ground their reasoning in evidence, plan over long horizons, call the right tools at the right time, and turn results into actionable scientific insight. You will work at the frontier of agentic AI for science, implementing code, running experiments, and owning results end-to-end.

Your Work May Involve

• Designing closed-loop agentic discovery systems that iterate between hypothesis, planning, tool use, observation, and refinement
• Engineering long-horizon agents: planning, memory, context management, and recovery across runs that span hours to days
• Building scientific tooling for agents (computational chemistry tools, simulation engines, property predictors, literature and patent search) and the protocols by which agents call them
• Implementing grounded reasoning over scientific literature, turning papers, patents, and databases into structured, citeable evidence that anchors agent hypotheses
• Building evaluation harnesses for agent reliability, factuality, and scientific value including designing benchmarks where simple QA metrics fall short
• Working with multimodal LLMs to parse molecular structures, figures, tables, and reaction schemes from scientific documents
• Designing and maintaining the molecular knowledge base and multimodal retrieval systems that grounds agent outputs
• Analyzing agent trajectories carefully, including debugging, failure analysis, and red‑teaming
• Communicating results clearly through plots, writeups, and paper‑ready figures

Profil Recherché

About You

• PhD in Machine Learning, NLP, or a related field or equivalent experience.
• A research track record in LLM agents, scientific ML, multimodal ML, or scientific NLP, including peer‑reviewed publications
• Strong implementation ability and comfort in research codebases (Python; PyTorch or JAX)
• Familiarity with the closed‑loop agentic discovery literature (e.g. ChemCrow, Coscientist) and current frontier work on long‑horizon agents
• Experience with LLM APIs, RAG, tool use, and agent frameworks (LangGraph, LlamaIndex, or similar)
• Strong communication and a bias toward clarity and honesty regarding results
• High agency: you push projects forward and take initiative

In Addition, The Following Would Be An Advantage

• Prior project, internship, or PhD work in drug or material discovery
• PhD work on LLM agents, tool‑augmented reasoning, multimodal scientific ML, or scientific knowledge bases
• Experience engineering long‑running agents (planning, memory, context management, failure recovery)
• Experience building scientific evaluation benchmarks for agentic systems