Inference Architect: MoE & Laneformer Research

Varsity Capital; Paris; Hybrid; EUR 90,000 - 120,000; Full time

Job summary

Kog is building a co-designed inference stack for real-time AI agents on standard datacenter GPUs. As a Research Engineer, you will focus on model architecture designed for inference, addressing communication, memory, parallelism, and execution constraints, with extensions to Laneformer 2B and DTP.

You will work on architecture choices for large MoE models, post-training methods, and experiments connecting hypotheses to measurable inference behavior.

Qualifications

• We seek researchers who design or materially change model mechanisms and connect architectural choices to inference-time behavior.

Responsibilities

• Design inference-focused model architectures across attention, routing, residual structures, and parallelism.
• Extend Laneformer 2B and DTP.
• Develop architecture choices for large MoE models and how routing/communication interact at inference time.
• Explore post-training methods that alter model structure or inference behavior.
• Experiment to link architectural hypotheses to measurable model and inference performance.
• Publish research results and translate ideas into the Kog inference stack.

Skills

Model architecture

MoE research

Transformer research

Inference optimization

Job description

Kog is building a co-designed inference stack for real-time AI agents on standard datacenter GPUs. As a Research Engineer, you will focus on model architecture designed for inference, addressing communication, memory, parallelism, and execution constraints, with extensions to Laneformer 2B and DTP.

You will work on architecture choices for large MoE models, post-training methods, and experiments connecting hypotheses to measurable inference behavior.