Research Engineer

Kog; Paris; Hybrid; EUR 90.000 - 140.000; Vollzeit

Benefits dieser Stelle

Remote-friendly

Paris office week (mandatory)

Travel & accommodation covered

Equity

Zusammenfassung

Kog is building the fastest LLM inference engine on datacenter GPUs, with a focus on architectural innovations and inference-speed optimizations. You will design architecture variants, explore DTP and MoE ideas, and own post-training pipelines for open-weight models, while scaling to large MoE configurations.

This role combines research and engineering to push the limits of speed and quality. You will publish findings, present at conferences, and help develop autonomous AI agents to run

Qualifikationen

• Demonstrated work on model architecture as a primary object of work (paper, repo or thesis).
• Ability to reason about model design and hardware together, especially Transformers and MoE.
• Experience with inference-aware variants (DTP, Ladder Residual, PT-Transformer) and post-training methods.
• PhD or top engineering school with architecture work counted is highly valued.

Aufgaben

• Design new model architecture variants with routing, attention, and MoE under execution constraints.
• Extend Laneformer thesis with inference-aware variants and scale findings at scale.
• Own post-training pipeline across fine-tuning, evaluation, and adaptation of open-weight models.
• Scale the stack to large MoE models, addressing routing, expert parallelism, and communication.
• Publish findings as research papers and present at top venues.
• Contribute to autonomous AI agents for architecture research and training experiments.

Kenntnisse

Architectural design

Transformers

MoE architectures

Hardware-aware design

Research publication

Ausbildung

Top engineering school or PhD

Jobbeschreibung

About Kog

Kog builds the fastest LLM inference engine on standard datacenter GPUs. Our Kog Inference Engine generates 3,000 output tokens per second per request on a single 8× AMD MI300X node and 2,100 on an 8× NVIDIA H200 node (FP16, batch size 1, no speculative decoding).

We co-design the model architecture and the execution engine together. Our Laneformer model uses Delayed Tensor Parallelism (DTP), a novel architecture that restructures the Transformer dependency graph so inter-GPU communication overlaps with computation rather than blocking it.

We pre-trained a 2B-parameter DTP model on 6T tokens on 256 H100 GPUs.

We are a team of 11 people, including 10 engineers and 5 PhDs.

Test it at playground.kog.ai. Read the technical details on the Kog Labs blog.

What you will work on

You will imagine, design, and run experiments to understand how architectural decisions propagate through inference behavior, morph existing open-weight models into architecture variants optimized for speed, and turn findings into measurable gains in generation speed and model quality.

• Design new model architecture variants, including routing strategies, attention mechanisms, and MoE structure, with execution constraints as a first-order design input.

• Extend the Laneformer thesis by exploring inference-aware architectural variants such as DTP, Ladder Residual, and PT-Transformer, and finding what compounds at scale.

• Own the post-training pipeline across fine-tuning, evaluation methodology, and adaptation of existing open-weight models toward architecture variants optimized for inference speed.

• Scale the stack to large MoE models such as DeepSeek v4 and Qwen 3, working through routing, expert parallelism, and communication patterns at inference time.

• Write up findings as research papers, submit them to top venues, and present them at conferences.

• Contribute to building AI agents that will perform architecture research and training experiments autonomously, starting from the research foundations we are building now.

What we look for

• You have designed or changed model architecture, where the structure itself was the object of the work. Showing that work, a paper, a repository, or a thesis, is a requirement to move forward.

• You reason about model design and hardware together, tracing how communication structure and layer dependencies shape inference behaviour, with fluency in Transformers and MoE deep enough to weigh trade‑offs.

• Stronger signals include inference‑aware architectural variants such as DTP, Ladder Residual, or PT‑Transformer, and post‑training methods such as fine‑tuning, preference optimisation, or quantisation, including at research scale.

• A top engineering school or a PhD with concrete architecture work counts, even without industry experience.

What we offer

• Direct access to AMD and NVIDIA datacenter GPUs from day one

• A team where creativity and technical judgement carry weight and where the people closest to the problem shape the key decisions

• Problems that sit on the critical path of model execution speed and that directly influence what the system can become

• A remote‑friendly working model, with one mandatory week per month in our Paris office. Travel and accommodation covered by the company.

• Compensation aligned with top AI research profiles, including equity