Research Engineer

Job description

About the Role

Scaling foundation models to structured data is not a solved problem. Unlike text or vision, tabular data has no canonical tokenization, no natural sequence, and no shared feature space across datasets, which means standard scaling laws, positional encodings, and pretraining objectives don't transfer. The architectural and optimization challenges are fundamentally different, and largely open.

At Neuralk, we're building the infrastructure and the models to solve this. As a Research Engineer, you'll work at the intersection of systems and research — owning the pretraining stack, pushing inference to production-grade latency, and writing the low-level GPU code that makes large-scale experimentation possible. The same people who build the systems shape the research direction.

Role & Responsibilities

Transformer inference optimization : Design and implement optimizations to reduce latency and improve scalability of our foundation models in production: attention mechanisms, KV-cache, batching strategies, quantization.

Pretraining optimization : Build and maintain the training infrastructure that enables us to scale pretraining efficiently (distributed training, mixed precision, gradient checkpointing, throughput optimization)

CUDA kernel development : Write low-level GPU kernels to optimize critical bottlenecks, from custom attention variants to data preprocessing pipelines.

Architecture experimentation : Collaborate with the research team to implement, train, and evaluate novel architectures adapted to structured data at scale.

Tooling & infrastructure : Develop the internal tooling (experiment tracking, evaluation pipelines, benchmarking) that enables rapid and reproducible research iterations.

Research contribution : Contribute to publications and engage with the broader ML research community.

Profile

Master's or PhD in Computer Science, Machine Learning, or a related field

5+ years of experience in ML systems, with a strong focus on training and inference optimization

Deep proficiency in Python and PyTorch, with strong software engineering practices

Hands-on experience writing CUDA kernels or using low-level GPU optimization libraries (Triton, cuBLAS, CUTLASS)

Solid understanding of transformer architectures, training dynamics, and scaling behavior

Experience with distributed training frameworks (DeepSpeed, FSDP, or equivalent) and cluster environments (SLURM)

Excellent communication skills in English

Self-starter, autonomous, and comfortable in a fast-paced startup environment

Bonus points

Publication record at top ML venues (NeurIPS, ICML, ICLR, MLSys)

Experience with inference serving frameworks (vLLM, TensorRT, ONNX)

Contributions to open-source ML infrastructure

Experience with structured or tabular data

What We Offer

A competitive salary

Equity (BSPCE), to reflect the value you bring to Neuralk and to foster a shared journey

Substantial compute resources and state-of-the-art ML infrastructure

Comprehensive health insurance

French level paid leave and time-off work

and more to come as we grow

! Image 1: Fantasy landscape with a towering floating mountain covered in greenery and vertical structures, surrounded by clouds, with a winding river in the foreground under a colorful sky.