Senior ML Engineer, Applied LLMs

About the role

Halcyon builds document understanding for regulated industries. Our customers cannot tolerate a confident wrong answer, which makes evaluation and retrieval quality the center of the engineering problem rather than an afterthought.

What you will do

Own the retrieval pipeline: chunking, hybrid search, reranking
Build the evaluation harness that decides whether a model change ships
Fine-tune open models where a hosted API is the wrong answer on cost or privacy
Bring inference cost per document down; it is currently the largest line in our infrastructure bill

What we look for

Shipped a retrieval-augmented system that real customers depend on
Opinionated about evaluation, and able to say why benchmarks mislead
Experience with parameter-efficient fine-tuning
Five or more years in production ML

What the assistant changed, and why

Change

Summary rewritten around retrieval quality and evaluation

Why

The posting mentions evaluation in three of four responsibilities. Your summary led with model training.

Change

Added the eval harness you built at your current employer

Why

It was a single sub-bullet. It is the closest match to their central problem and now leads that role.

Change

Quantified the LoRA fine-tuning work with the cost reduction

Why

The posting names inference cost explicitly; your 60% reduction was in the resume without the number attached.

Change

Reordered skills to put retrieval, evaluation and PEFT first

Why

Their screen is likely keyword-filtered, and these were below the fold.

What matched

RAG system in production with paying customers
Evaluation harness built and owned
LoRA and QLoRA fine-tuning experience
Six years of production ML

What it would not stretch

No regulated-industry experience. Not claimed — worth preparing an answer for.

Awaiting your approval

The draft is written and the form is filled in on Halcyon Labs’s own site. It sits here until you approve this one application. There is no approve-all, and nothing is submitted without you.