Senior Machine Learning Engineer

We’re looking for a Machine Learning Engineer who can own the full lifecycle of alarge language model in production - from training and fine-tuning throughdeployment at scale and rigorous evaluation of what actually ships. This isn’t aresearch-only role and it isn’t a pure infrastructure role: you’ll need to begenuinely fluent in all three, because the hardest problems in this space live atthe seams between them - a training decision that breaks serving latency, adeployment optimization that silently degrades output quality, an eval result thatdoesn’t predict real-world behavior.

What you’ll do:

Training & fine-tuning

• • Design and execute pretraining, continued pretraining, and fine-tuning runs (SFT, DPO/RLHF-style alignment, LoRA/QLoRA and full-parameter approaches) against clear, measurable objectives
• • Own data pipeline decisions that materially affect model quality - curation, deduplication, mixture weighting, and contamination checks against eval sets- Run and interpret distributed training (multi-GPU, multi-node) using frameworks such as FSDP, DeepSpeed, or Megatron-style parallelism, and diagnose failures that only show up at scale (loss spikes, stragglers, checkpoint corruption)
• • Make and defend real tradeoffs between model size, training cost, and downstream performance

Deployment

• Take a trained model to production: quantization, batching strategy, KV-cache management, and serving framework selection (e.g. vLLM, TensorRT-LLM, TGI) with explicit latency/throughput/cost targets, not just “make it run”
• Design for the failure modes specific to LLM serving - tail latency under load, graceful degradation, prompt injection surface area, and safe fallback behavior
• Build the operational muscle around this: monitoring, alerting, and rollback paths for a model in production, treated with the same rigor as any other critical service, not as a one-off notebook export

Assessment & evaluation

• Build and maintain evaluation harnesses that go beyond running published benchmarks - including task-specific eval sets that reflect your actual product’s use cases, not just leaderboard performance
• Design human evaluation protocols where automated metrics fall short, and know which is which
• Own regression detection: catching quality drops introduced by a new checkpoint, a prompt template change, or a serving optimization before they reach users- Contribute to safety and robustness evaluation - hallucination rate, adversarial/ red-team testing, and behavior under distribution shift - as a first-class part of the release process, not an afterthought

• 4+ years of applied ML engineering experience, with at least **2 years working directly on large language models in a production context
• Real production deployment experience - you’ve shipped a model that served live traffic, and you can talk about the latency/cost/quality tradeoffs you made- Strong software engineering fundamentals
• Fluency with the modern LLM tooling landscape (training frameworks, serving frameworks, eval tooling)
• Comfort with ambiguity - you’ll be asked to define what “good” means for a model behavior that doesn’t have an established benchmark