R-03 · RACD

RESEARCH

Founding Machine Learning Engineer, RACD

The distance between a model that works in a notebook and a model that holds at three in the morning is where this role lives.

FULL-TIME · FOUNDING · REMOTE · OVERLAPPING HOURS

Apply for this role

OPENINGS

1

One. When it is filled, this page comes down.

BASE SALARY

£230,000 to £390,000

The band, published. The offer is the same whether or not you push for more.

EQUITY

0.55% to 1.10% of RACD

Founding equity in the company you would actually work on.

01

The problem this role exists for

RACD's research output has to run continuously over heterogeneous, high-cardinality streams, on data that arrives late, out of order and occasionally not at all. You would own training, evaluation, serving, and the drift detection that tells us when yesterday's model stopped describing today's estate.

This is not a role that receives models over a wall. You will be in the research conversation early enough to say that a formulation is beautiful and unservable, which counts here as a contribution rather than an obstruction.

More on the company itself: RACD

Correlated detection versus threshold alertingThree independent signals drift at different times. Correlation identifies the shared cause early; conventional threshold alerting fires only after user-visible impact.SENSOR · VIBRATIONQUEUE · DEPTHSERVICE · p99 LATENCYRACD correlates → alarmthreshold firesLEAD TIME · THE ENTIRE PRODUCTDIVERGENCE BEGINSIMPACT
Figure 1: Correlated detection against threshold alerting. Three unrelated-looking signals diverge from one cause. The gap between correlation and the conventional alarm is the interval in which an outage is still preventable, and closing it is the job.

02

Your first week, and a normal Tuesday

Your first fortnight: get the current best model onto live traffic in a shadow deployment, running alongside the threshold alerting a customer already has, and produce the first honest side-by-side of the two. That comparison does not exist yet, and almost everything downstream of it gets easier once it does.

A normal Tuesday looks closer to systems engineering than to modelling. The evaluation harness, a latency regression somebody introduced yesterday, a conversation with a researcher about whether a formulation can be served at all, and the slow work of turning every alarm an engineer acted on or dismissed into a label. You are on call for what you ship, on a rota designed by the person in R-07.

WHAT YOU WOULD BE WORKING IN

Python for training and evaluation, something compiled on the serving path, a streaming layer, a columnar store, and Kubernetes underneath all of it. Broad on purpose: we care that you can get a model to hold under load, not which framework you got it there with.

We do not screen on tools. Nothing above is a requirement, and a stack list is a poor proxy for whether somebody can do the work. If you have gone deep in something else and would arrive with a view about why half of this is the wrong choice, that is an argument we would like to have.

03

What you would own

Not a ticket queue and not a slice of somebody else’s roadmap. Four things with your name against them in the decision log, from the first fortnight.

01

Serving

Inference over live streams, inside a latency budget an on-call engineer would accept.

02

Evaluation

The harness that decides whether a change is an improvement, and that nobody argues with.

03

Drift

Knowing before the customer does that the model stopped being true.

04

The loop

Every alarm acted on or dismissed is a label. You close that loop.

04

What we are looking for

MINIMUM

What we would need to see

  • You have put a model into production and kept it there, and you can say what broke and what you changed
  • Comfort with streaming or out-of-order data and the failure modes it introduces
  • You can read the paper the research team is arguing about and hold a view on it
  • You have owned the operational consequences of something you shipped

PREFERRED

What would move you up the pile

  • Anomaly detection, forecasting or observability tooling in a previous life
  • An opinion about why most feature stores are the wrong shape
  • Shadow deployments, online evaluation or interleaving in production
  • You have built the labelling loop as well as the model that consumes it

Nothing in the preferred column is a filter, and neither list is applied before a human reads you. No degree is required for anything on this page, and we have hired people with doctorates and people with no degree at all into the same rooms. If you meet most of the minimum and not all of it, write anyway and say in the first paragraph which part you do not meet.

05

Pay, and everything around it

£230,000 to £390,000 base, plus 0.55% to 1.10% of RACD, reviewed upward annually without you having to ask. The band is published here rather than discovered three conversations in, and the offer is the same whether or not you negotiate, because the band is the band.

Everything else is what the whole group gets and is not negotiated per offer: a holiday floor rather than a ceiling, health and dental and vision for your family, sixteen to twenty weeks of parental leave for every parent, a funded setup, monthly paid group meetups somewhere in the world, one required funded trip abroad each year, a protected day a week on the research ventures, and no non-competes.

The full list, the salary band chart and the visa sponsorship map are on the careers page. We sponsor without qualification across seventeen jurisdictions and publish exactly which.

06

Apply for this role

Four steps: a conversation with somebody who would actually work with you, one real problem drawn from work we are genuinely doing, a working session on your solution, and a decision within a week of that session. The problem is paid if it runs long, and you keep the work either way.

Or write to careers@cnsolarlabs.com.

What you send here reaches us and no one else. We use it to reply to you, we do not add you to any list, and you can ask us to delete it at any time. See the privacy policy.