Senior LLM Engineer

Multiverse Computing

• Train and post-train LLMs and multimodal models on multi-node GPU clusters using SFT, preference optimisation, RL (e.g. GRPO), and on-policy and offline distillation.

• Design compression techniques (quantization, pruning, quantization-aware training, quantum-inspired tensor-network methods) and healing recipes that recover the quality of compressed models.

• Build RL environments, reward functions and verifiers for reasoning, coding and agentic tasks, and extend RL methods to world models and Physical AI.

• Research and prototype post-transformer and hybrid architectures (state-space models, linear attention, MoE), from small-scale ablations to scaling recipes.

• Build data pipelines for curation, filtering, deduplication and large-scale synthetic data generation.

• Design rigorous ablations and evaluations on current benchmark suites (reasoning, coding, long-context, instruction following, agentic) as well as internal evals.

• Port and optimise models for inference on various accelerators and edge chips, profiling latency, memory and throughput.

• Write up experiments and results as internal reports and, where relevant, papers and open-source releases.

• Review code, mentor teammates, and follow and bring in the latest research.

• Master's or PhD in Computer Science, AI, Mathematics, Physics or a related field, or equivalent industry experience.

• Senior: 4+ years of hands-on experience training or post-training transformer models.

• Strong PyTorch. You are comfortable working below library level: model code, custom losses, parallelism configuration.

• Experience with distributed training (FSDP, DeepSpeed or Megatron-LM; data, tensor, pipeline and expert parallelism) on multi-node clusters with Slurm/K8s.

• Hands-on post-training experience: SFT, DPO, and at least one RL method (PPO, GRPO or similar).

• Solid mathematical foundations in linear algebra, optimisation and probability, and an understanding of training dynamics.

• Good understanding of GPU fundamentals: memory hierarchy, mixed precision, profiling, throughput.

• Rigorous experimental practice: ablation design, reproducibility, experiment tracking, noise-aware evaluation.

• Strong debugging and communication skills. Fluent English.

• Experience with the NVIDIA stack (Megatron-LM, Megatron Bridge, NeMo RL, NeMo Gym, NeMo Curator) or similar frameworks (verl, TRL, OpenRLHF).

• RL infrastructure: rollout engines (vLLM, SGLang), asynchronous RL, reward model and verifier design, agentic environments.

• Model compression: post-training quantization and QAT (GPTQ, AWQ, FP8/FP4), pruning, distillation, low-rank or tensor-network methods.

• Post-transformer architectures: SSMs/Mamba, linear attention, hybrid models, diffusion LMs.

• Physical AI: vision-language-action models, world models (e.g. JEPA-style or video world models), robotics simulation.

• Multimodal model training (vision-language, audio).

• Inference on alternative hardware: ExecuTorch, llama.cpp, ONNX Runtime, OpenVINO, vendor NPU SDKs, custom kernels (Triton, CUDA).

• Large-scale data processing and synthetic data generation.

• Evaluation tooling (lm-evaluation-harness, lighteval, NeMo Evaluator) and familiarity with Artificial Analysis.

• Publications at top venues (NeurIPS, ICML, ICLR, ACL, etc.), notable open-source contributions, or models released on Hugging Face.

• Spanish.