Hélain Zimmermann
Co-Founder & CTO @ Ailog
Stockholm, Sweden
I ship AI systems: RAG, agents, fine-tuning. Most of what I write here comes from code that actually runs in production.
Right now I am finishing an MSc in Machine Learning at KTH Stockholm, alongside an engineering degree at ENSIMAG Grenoble. Before Sweden I spent a summer at INRIA Grenoble working on how language models memorize personal data when fine-tuned on sensitive corpora. That work became arXiv:2501.02407, which I co-authored.
In March 2025 I left the Nsigma Junior-Enterprise and co-founded Ailog. We build AI features for small and mid-sized teams: retrieval over whatever internal docs they already have, agent workflows that replace brittle scripts, automation glue between tools they already pay for. Constraints: GDPR when it applies, teams of 3 to 20, tight budgets, and a demo that cannot survive a Monday morning is useless.
What I work on
Ailog: consulting and SaaS
RAG over whatever docs clients already have (PDFs, Notion dumps, ticket history). Agent orchestration for internal workflows. Glue between tools they already pay for. My job is to scope the problem, pick boring infrastructure when it fits, and ship something ops can actually run.
Research: parameter-efficient fine-tuning
Multi-expert LoRA with a learned router, in the spirit of Mixture-of-Experts but at the adapter level. The question: given 8 tasks (SQuAD, IMDB, CoNLL-2003, WikiText-2, GSM8K, XSum, CommonsenseQA, MNLI), does a gating network learn to send each task to the right expert without manual partitioning? Phi-2 (2.7B) is done across 7 configurations. Top-K sparse routing with k=2 shows the strongest specialization so far (task specialization score 0.0657, versus 0.0065 for a load-balanced baseline). Qwen2.5-0.5B is currently training. Llama-3.2-3B and Gemma-2-2B are queued for the same 7-config sweep.
Research: validation of LLM-based social simulations
SimValid is a multi-scale validation framework (micro, meso, macro) for social simulations driven by LLM agents. It is built on an open-source rewrite of MiroFish where Zep Cloud is replaced by a local NetworkX memory graph. 15 quantitative metrics, 5 per scale. Methodological goal: most published LLM simulations claim emergent behavior with no benchmarks. I want to show you can actually test those claims. Target venue: AAMAS 2027.
A broader list of projects (open-source, academic, personal) lives on the homepage.
Research
Towards the Anonymization of the Language Modeling
arXiv:2501.02407 · 2025 · INRIA Grenoble
With Antoine Boutet, Lucas Magnana, Juliette Sénéchal
Language models fine-tuned on sensitive corpora memorize personal data and leak it under targeted prompts. We tested two training schemes against this: a masked objective for BERT-style models, a causal objective for GPT-style ones. Both target direct identifiers (names, numbers) and indirect ones (contextual hints that re-identify a person). We evaluated on a medical dataset against several baselines. Privacy is preserved; utility drops less than I expected. Numbers and code are on the arXiv page.
Experience
Mar 2025 – Present
Co-Founder & CTO
Consulting and SaaS in AI, automation, and applied mathematics for European companies. I lead engineering: RAG systems, LLM agents, privacy-aware pipelines.
Mar 2025 – Present
Technical Consultant — RAG
Nsigma Junior-Enterprise • Grenoble
RAG systems for student-run consulting projects. Previously Technical Director (Nov 2023 – Mar 2025).
Jun – Sep 2024
Research Intern
Privacy-preserving NLP. Co-authored arXiv:2501.02407 on text anonymization and memorization risks in language models across French and English.
May – Jun 2023
IT Technician Intern
IE-Concept
IoT and embedded systems.
Education
2025 – 2026
MSc Machine Learning
KTH Royal Institute of Technology • Stockholm
Joint cursus with ENSIMAG
2023 – 2026
Engineering Degree
ENSIMAG - Grenoble INP • Grenoble
Applied Mathematics & Computer Science
2021 – 2023
Cycle Préparatoire Polytechnique (CPP)
Grenoble INP • Valence
Integrated engineering preparatory cycle
Get in touch
Consulting work (RAG, agents, AI engineering) goes through Ailog, or email me directly if you prefer. For editorial feedback on an article, same email, just mention which one.
Editorial policy and corrections → /editorial-policy