Software Engineer, Transformers

About the role

Our products and client work run on large language models, and the gap between a good result and a great one is often in the model itself: how it was adapted, how it is evaluated and how it is served. As a Software Engineer focused on transformers, you will work across that whole stack, from reading a new architecture paper to shipping a fine-tuned model that answers in under a second.

You will join a small group that sits between research and product. Some weeks you will run careful experiments; others you will profile an inference server and cut its cost in half. Both matter equally here.

What you will do

• Fine-tune, distil and evaluate open weight transformer models for the tasks our products and clients depend on, such as code generation, structured output and retrieval
• Design and run experiments on architecture, data mixtures, training objectives and optimizers, and write up what you learn so others can build on it
• Build and maintain training and evaluation pipelines that are reproducible, efficient and easy for the team to use
• Optimise inference for latency and cost with quantization, batching, speculative decoding and KV cache management
• Build evaluation sets and harnesses that tell us honestly whether a model got better, and catch regressions before they reach people
• Work with product engineers to put models into real features, and bring what you learn in production back into training

You may be a good fit if you

• Have 5+ years of software engineering experience, with at least 2 years working hands on with deep learning models
• Are fluent in Python and PyTorch, and comfortable reading and changing model code at the level of attention layers and loss functions
• Understand transformer architectures deeply: attention and its variants, positional encodings, normalisation, mixture of experts and how they trade off
• Have trained or fine-tuned models across multiple GPUs and debugged the problems that come with it
• Design experiments carefully, measure honestly and are comfortable saying a result did not hold up
• Write clear code and clear notes, and enjoy working with people from other disciplines

Strong candidates may also have experience with

• Distributed training with DeepSpeed, FSDP or Megatron
• Inference servers such as vLLM, TensorRT-LLM or SGLang
• Writing or tuning GPU kernels in CUDA or Triton
• Reinforcement learning from human or AI feedback, or preference tuning such as DPO
• Published research or well known open source work in machine learning

About NerfedLab

NerfedLab builds AI products, helps businesses put AI to work in their everyday operations, and runs a research team that publishes its work in the open. Zocri, our product, lets anyone make an app or a website by describing it, then run the business behind it: sign-in, payments, messages and marketing in one place.