Research Scientist
Save this job and keep your search organized
Create a free account to save jobs, create alerts and return to this listing from your dashboard.
By continuing, you agree to our Terms & Privacy Policy.
About the Role
OceanLum turns media captured by divers and snorkellers into a verifiable record of which individual marine animals were encountered, and where. Those identifications come from many sources — guests, crew, annotators, our own models — that disagree, that are inconsistent with themselves, and a few of which have reason to claim an encounter that never happened. No signature can bind a claim to a living animal in open water, so the problem is statistical, not cryptographic.
You will design and validate a non-cryptographic Byzantine fault tolerant analytic for an organic identification labelling system : an estimator that recovers a defensible consensus identity, with honest confidence bounds, from reporters of unknown and variable reliability. This is a research post. You own the question, the method, and the evidence that it works. We expect the result to be publishable as well as shippable, and we support publication.
In this role, you will
Specify the fault model — the failure and adversary classes for identification claims, and the fraction of faulty reporters the analytic must tolerate.
Design and pre-register the consensus estimator: reliability-weighted latent-state inference over multi-source claims.
Build detection for anomalous and fabricated reporting.
Characterise the real error rate — sensitivity, specificity, operating threshold — by simulation and a field validation study with our operators.
Establish where consensus fails and what confidence a published label can honestly carry.
Hard requirements. We cannot consider applications that miss any of them.
Hold a PhD in a STEM discipline , conferred at the time of application. Field is open — statistics, psychometrics, epidemiology, clinical research, machine learning, quantitative ecology.
Have run large-scale studies on unreliable real-world data , where ground truth was unavailable and the instrument was itself a source of error.
Know the statistics of unreliable measurement: inter-rater reliability (ICC, kappa), latent-variable or IRT modelling, signal detection theory, principled handling of missing data.
Have a peer-reviewed publication record as a substantive author, and write analysis plans that survive outside scrutiny.
Be fluent in Python or R , and write analysis code others can reproduce.
Be legally permitted to work in Hong Kong for the full engagement period.
Be available full-time and on-site for a minimum of six months.
Nice to have
Turning a research method into a product or a regulated decision system.
Commercial marine, aquaculture or wildlife data work; fieldwork with non-technical participants.
Robust statistics or distributed consensus literature; computer vision for individual animal re-identification.
Comfort explaining an error bound to a customer or a regulator without hiding behind it.
About OceanLum
OceanLum is the platform for marine excursion operators. Our engine recognises individual marine animals by their biometric markers — spot patterns, scarring, fin shapes — not just their species, and turns guest media into encounter records operators, guests and researchers can rely on.