StellenbeschreibungppGiotto.ai is a Switzerland-based AI company building intelligence systems for Switzerland and Europe. /p pOur mission is to enable governments and enterprises to retain full control over the AI systems they use, without compromising access to the most advanced reasoning capabilities. From this control comes what matters most: protected data, preserved autonomy, and lasting strategic independence. /p pGiotto is a portable, configurable model and AI operating system with advanced reasoning capabilities, combining open and proprietary weights, datasets, and tools to deliver high performance, adaptability, robustness, and multi‑agency support. /p h3About the role /h3 pWe are looking for an bAI Research Scientist /b to help design, train, evaluate, and improve advanced AI systems. This role sits at the intersection of deep learning research, applied machine learning, and scalable engineering. /p pYou will work on research problems involving large language models, multimodal reasoning, synthetic data generation, model evaluation, representation learning, and task‑specific adaptation. You will be expected to move from ambiguous research questions to concrete hypotheses, experiments, prototypes, and eventually production‑ready methods in collaboration with engineering and product teams. /p pThis is a hands‑on research role: you will read papers, design experiments, train and fine‑tune models, build evaluation pipelines, analyze failures, and help translate research insights into reliable AI capabilities. /p h3Responsibilities /h3 pAs an bAI Research Scientist /b, you will: /p ul liDefine and execute research projects around LLMs, reasoning, multimodal models, synthetic data, and model adaptation. /li liDesign experiments to test hypotheses, compare architectures, evaluate training strategies, and measure model behaviour. /li liTrain, fine‑tune, and evaluate Transformer‑based models using modern deep learning frameworks. /li liWork on supervised fine‑tuning, preference optimisation, LoRA/PEFT methods, distillation, data augmentation, and evaluation‑driven model improvement. /li liBuild robust benchmarks and diagnostic evaluations for reasoning, generalisation, reliability, and task‑specific performance. /li liAnalyse model failures and propose improvements at the level of data, architecture, training objective, prompting, or inference strategy. /li liCollaborate with ML engineers to scale experiments across GPUs and distributed infrastructure. /li liContribute clean, reproducible research code, experiment configs, documentation, and internal reports. /li liStay up‑to‑date with relevant AI research and translate promising ideas into practical experiments. /li liHelp shape the company's research roadmap and identify high‑impact technical directions. /li /ul h3Required experience /h3 pWe are looking for someone with strong experience in several of the following areas: /p ul liDeep learning, especially Transformer architectures and modern sequence models. /li liLLM training, fine‑tuning, evaluation, or inference. /li liStrong practical experience with bPython /b and bPyTorch /b. /li liExperience with the Hugging Face ecosystem: transformers, datasets, tokenizers, model checkpoints, and generation APIs. /li liUnderstanding of training dynamics, optimisation, loss functions, overfitting, regularisation, and evaluation methodology. /li liAbility to design rigorous experiments and interpret results beyond headline metrics. /li liExperience working with large datasets, pre‑processing pipelines, and reproducible ML workflows. /li liStrong mathematical foundations in linear algebra, probability, statistics, and optimisation. /li liAbility to read research papers and turn them into working prototypes. /li liClear communication skills and the ability to explain research trade‑offs to technical and non‑technical stakeholders. /li /ul h3Relevant tooling and stack /h3 pThe role should stay close to the current ML engineering stack while adding research‑oriented tools. /p pExpected core stack: /p ul liPython /li liPyTorch /li liHugging Face Transformers / Datasets /li liCUDA‑aware GPU training /li liMLflow or Weights Biases for experiment tracking /li liDocker /li liRay or similar tools for distributed workloads and experiment orchestration /li liPyTorch Distributed, FSDP, DeepSpeed, or Accelerate /li liPEFT / LoRA / QLoRA /li livLLM /li lipytesbt /b and reproducibility tooling for research code quality /li /ul h3Nice‑to‑have tooling /h3 ul liTriton or custom CUDA kernels /li liGCS /li liRAG pipelines, vector databases, and embedding evaluation /li liData annotation, synthetic data generation, and human‑evaluation pipelines /li /ul h3Nice‑to‑have research areas /h3 pExperience in one or more of the following would be especially valuable: /p ul liLLM reasoning and planning /li liProgram synthesis or structured prediction /li liModel merging, distillation, and compression /li liReinforcement learning or preference optimisation /li liEvaluation of reasoning and generalisation /li liMechanistic interpretability or model analysis /li liRetrieval‑augmented generation /li liAgentic systems and tool‑using models /li liDistributed training at scale /li liLow‑level inference optimisation /li /ul h3Profile we are looking for /h3 pYou may be a good fit if you: /p ul liEnjoy turning unclear research questions into measurable experiments. /li liAre comfortable with both theory and implementation. /li liCare about reproducibility, clean experiment tracking, and honest evaluation. /li liCan move quickly from paper to prototype. /li liAre pragmatic: you know when to pursue a research idea deeply and when to stop. /li liLike working closely with engineers to make research usable in real systems. /li liAre excited by frontier AI problems but grounded in measurable progress. /li /ul h3Location Work Style /h3 pWe offer a full‑time bemployment /b in Switzerland. /p pHybrid model: /p ul liRemote work fully supported /li liTeam gathers bone week per month /b in the Swiss office /li /ul pExceptional candidates residing elsewhere in bEurope /b may be considered. /p /p #J-18808-Ljbffr
