Senior Generative AI Research Engineer — Sustainable Cosmetics & Pharma R&D (Tokyo · Paris · Remote-friendly)

Context

A Tokyo-listed Japanese pharma & cosmetics group (¥300B+ revenue, multi-decade brand portfolio in eye care, skincare and functional cosmetics) is launching its flagship AI program. This is a generative AI / foundation models research role — not a classical data science assignment.

The client's CIO — 26 years at IBM — set an explicit course: replicate in-house what the most ambitious pharma × AI partnerships delivered in 2024–2025, pushing three axes: fully leveraged generative AI (not statistical ML wrapped in dashboards), unsupervised learning (generate formulations and discover categories without historical data), and sustainable cosmetics by design (biodegradability, sourcing impact, multi-geo regulation baked in from day 1).

Mission

You will build, with Abbeal and the client's R&D, a custom foundation model for cosmetic & pharmaceutical formulation. The architecture combines:

• Formula Generator — transformer / diffusion / GNN architectures on molecular graphs and ingredient embeddings, trained on a curated corpus of 10,000+ cosmetic ingredients.
• Multiple Formula Evaluators — parallel scoring heads for regulatory compliance (multi-geo), sustainability (carbon, sourcing, biodegradability), cost and product performance.
• Self-supervised category discovery — unsupervised representation learning to surface new product categories.
• External data pipelines — RAG on regulatory streams, scientific literature, supplier specs.
• Decision dashboards — explainable outputs for R&D scientists and brand leadership.

Inspiration: molecular foundation models in the lineage of MoLFormer, ChemBERTa and RXN for Chemistry, adapted to cosmetic formulation.

Roadmap

• Phase 0 — Audit (2–3 weeks): use-case mapping, data due diligence, current R&D stack assessment.
• Phase 1 — PoC (3 months): MVP foundation model, sustainability scoring head, evaluation harness on 2 product categories.
• Phase 2 — Production (12+ months): scaling, R&D team integration, production deployment, governance.

Logistics

• Work mode: remote-friendly from Paris or Tokyo, with regular Tokyo travel (~1 week/month on-site Tokyo minimum during Phase 0–1).
• Status: Freelance via Abbeal or CDI (Abbeal Japan contract with international mobility package via Mobbeal for candidates open to relocation).
• Working language: English. Japanese (JLPT N3+) is a real plus, not a prerequisite.
• Start: Q3 2026 — audit phase can start as soon as profile is validated.

Required profile — mandatory

• 5+ years building generative AI / foundation models applied to a scientific domain (chemistry, materials, drug discovery, life sciences).
• Hands-on transformer / diffusion / VAE / GNN on molecular representations (SMILES, SELFIES, 3D graphs).
• Mastery of state-of-the-art unsupervised & self-supervised learning, beyond LLM fine-tuning.
• Scientific background — PhD in computational chemistry, ML, materials science or computer science with strong domain experience, or industrial equivalent.
• Fluent business English — you will debate foundation model architecture with an ex-Distinguished Engineer level technical sponsor.
• Ability to scope and execute a PoC → production roadmap over 12–18 months.

Valued strengths

• Conversational or professional Japanese (JLPT N3+).
• Experience at IBM Research, ETH Zurich, EPFL, Institut Pasteur, Institut Curie, Imperial College or equivalent.
• Direct experience in cosmetic, fragrance or food formulation R&D (Tier-1 cosmetics, Givaudan, IFF, Firmenich, Iktos, Prose, etc.).
• Experience with Japanese clients and indirect communication culture.
• Publications (NeurIPS, ICML, JCIM, JACS, Nature ML) or open-source contributions.

Tech stack

• Modeling: PyTorch, JAX, Hugging Face Transformers, RDKit, DeepChem, PyTorch Geometric.
• Cloud: AWS or GCP (Tokyo region likely).
• Ops: MLflow, Weights & Biases, DVC, Airflow / Prefect.
• Data: PostgreSQL + vector store (Pinecone, Weaviate, pgvector), Snowflake or BigQuery.
• Evaluation: SHAP, integrated gradients, uncertainty quantification, A/B testing.

Why join us

• Unprecedented program for a major Japanese R&D player — same lineage as the foundation models / sustainable cosmetics partnerships of 2025.
• Frontier research, applied — you own the architecture choices for a flagship multi-year model.
• Demanding technical sponsorship — the CIO, 26 years at IBM, reads your papers.
• Modern stack, no legacy ML debt — green-field architecture.
• International setup — French / Japanese / global engineering culture, follow-the-sun Paris, Montréal, Tokyo.
• Abbeal mobility option — visa & relocation support via Mobbeal to move to Tokyo.

Apply

If this profile fits you — or someone in your network — let's talk. The CIO sponsor set a mid-June 2026 deadline for technical alignment: it moves fast. Write to sebastien@abbeal.com with subject line "GenAI Foundation Models Cosmetics Tokyo".