Kestrel · Remote (EU) · Posted
Applied ML Engineer, Ranking
- €120k–€158k
- Single-node
- Applied ML & product
- Remote · Full-time
Ranking here is the product surface most people see first, and the difference between a good model and a shipped improvement is mostly everything around the model. Feature freshness, training-serving skew, and an experiment design that can actually detect the effect size you care about.
You would own that whole loop: offline model work in PyTorch, features through Feast, and the serving path in Go. You will spend more time than you expect on why offline gains did not reproduce online, which is the most interesting part of the job.
We want someone who is suspicious of their own offline metrics.
What we’re looking for
- Shipped a ranking or recommendation model to real users
- Have diagnosed a training-serving skew problem
- Can design and read an online experiment without hand-waving the statistics
- Comfortable owning a service, not only a notebook
Tell us about a model that looked better offline and did not hold up online. What was the cause, and what would you check first next time?
That is the application. No cover letter, no screener, no timed test — Anna reads your answer.
You would report to Marek Nowicki, Engineering Manager, Discovery.