001 / depth over breadth
One thing, well.
The lab has five research threads and four engineering tracks. Everyone owns one. We resist the urge to spread thin.
/ careers at darkfield
Build an AI that
Build an AI that
writes its own pipelines.
We're a small team in the UK, hiring deliberately. The work splits between research on the two models and the orchestration layer that connects them, and applied engineering that puts both into industrial deployments. Below: the four roles open right now, and how to get in.
/ how we work
Four things to know
Four things to know
before you write to us.
002 / publish
Open by default.
Our research is written up as it goes, held internal during the private beta, and published once it closes. The parts we don't publish — by contract, not by ideology — are the customer pipelines and the per-camera adapters.
003 / read the hardware
Vision is a systems problem.
Every researcher here can read a kernel trace. Every engineer can read a paper. We don't have a wall between research and infra, and we like it that way.
004 / no algorithmic gauntlet
Take-homes, not whiteboards.
Hiring is a long-form take-home (paid) and two panels. We read what you've shipped and what you've written. We do not optimize for performance under fluorescent lights.
/ open positions
Four roles open today.
If your fit isn't here, write anyway. We keep a slow inbox and we read it.
-
Train new versions of SEE on industrial corpora. Push open-vocabulary segmentation past the SAM-3 frontier. Author research that we'll publish once the beta closes.
What you'll do
- Lead training runs on the lab's H100 cluster
- Co-design the next iteration of SEE's tokenizer
- Author or co-author 1–2 internal write-ups per year, slated for publication post-beta
What we look for
- PhD or equivalent rigour in vision / ML
- Track record of public artifacts before joining — papers, code, weights
- Comfort with low-level training infra (DDP, NCCL, Triton)
-
Make THINK better at supervising SEE. The work spans evaluation, planning, tool-use, and the long-horizon question of when a model should retrain itself.
What you'll do
- Design self-evaluation protocols THINK can run unattended
- Define how THINK invokes SEE — the typed surface between planner and perceiver
- Lead one of the lab's research write-ups, queued for publication after the beta
What we look for
- Background in agents, RL, or LLM tool-use
- A taste for adversarial evaluation
- Comfort with both natural language and pixel space
-
Take a partner from "we have a question" to "the model is running and grading itself." You'll spend time on factory floors and warehouses; you'll write the per-camera adapters that make SEE specialize on day one.
What you'll do
- Onboard new partners end-to-end (3–5 per year)
- Build and run per-camera finetune jobs
- Feed real-world failures back to the foundations group
What we look for
- Production CV experience — the messier the better
- Comfort on a factory floor at 6am
- Excellent technical writing — partners read everything
-
Own the training cluster end to end — schedulers, file systems, bandwidth. SEE has to retrain often and per-camera, which means the infra has to be cheap and fast at small jobs as well as big ones.
What you'll do
- Run the lab's training cluster end-to-end
- Reduce per-camera finetune cost by 4× this year
- Build the rolling-eval pipeline THINK uses to grade SEE
What we look for
- Linux internals, NCCL, slurm or k8s, Triton
- Track record on multi-node training
- Bias for measurement over conjecture
/ the process
Two weeks, end to end.
Two weeks, end to end.
No surprises.
step 01 · day 0
Application
A short email. Tell us which role and link to one or two artifacts you're proud of. No CV required, but welcome.
step 02 · day 1–6
Take-home
Paid. Five days, ~10 hours of focused work. The brief is the same kind of problem you'd see in your first month here.
step 03 · day 7–10
Two panels
A technical conversation about the take-home, then a longer one about how you think and how you'd work with the team. Both with people you'd actually work alongside.
step 04 · day 11–14
Decision + offer
A written decision either way, with reasoning. If it's an offer, top-of-band, equity, no surprises.
/ apply
A single inbox. A real person reads it.
Send a short note to hello@linox.co.uk. Mention the role, link to anything you're proud of, and tell us why now. We try to reply within five working days.