Role: Data Scientist
Contract: Permanent
Salary: Very Competitive
Location: Oxford / London
Security clearance: This role requires eligibility for UK security clearance (BPSS and SC). SC requires 5 years of continuous UK residency. An active SC or DV clearance is a strong advantage. If you are not currently eligible, we will not be able to progress with your application.
A note from the Founders
Oxford Dynamics is at an inflection point.
We operate in some of the most complex and high-stakes environments in the world defence - national security, AI and robotics. The decisions we make now will define not just how fast we grow, but who we become.
You will work closely with the whole team. You will be trusted with judgment calls. You will influence the business. And you will see the impact of your work every day.
If you are excited by ownership, pace and purpose, and by building something that genuinely matters - we would love to hear from you.
Core Remit
You turn messy, real-world sensor and signals data into detection capability that tells an operator what matters - whether that calls for a machine-learning model, a deterministic rule, or a technique no one has tried yet - and you are honest about what the data can and cannot say.
Your Brief
- Build detection capability on real, hard data: electronic intelligence (signals from emitters), maritime vessel-tracking data, and fused multi-source feeds. The kind of data that arrives with gaps, errors, and deliberate manipulation.
- Pick the right technique for the problem - not the fashionable one. Sometimes that’s a machine-learning model; often it’s a deterministic rule, a statistical test, or a physics- or behaviour-based heuristic; sometimes it’s something you invent. We care that it’s correct, explainable, and defensible, not that it counts as “AI”. Novel thinking and new techniques are actively encouraged.
- Engineer the features and signals that make or break detection - the per-event aggregates, rolling-window statistics, and geospatial and temporal patterns that determine whether a rare anomaly is caught or missed.
- Define what “good” looks like for a given intelligence type (an INT) and the threat it serves, then design the techniques that enforce that quality and detect and control drift over time - so a model or rule that was right last month doesn’t quietly rot as the world changes.
- Own the pipelines that get detection from idea to production fast. The goal is to ship and update models and rules at a rate of change that keeps pace with agentic workflows - continuous, automated, and safe, not a release every quarter.
- Own evaluation end to end: design the holdout, read the precision-recall trade-off, calibrate thresholds against how many false alarms an operator can actually tolerate, and say plainly where the signal is real and where it is noise.
- Take exploratory analysis and harden it into scored, reproducible capability that operators and our platform can rely on - moving it from “interesting notebook” to “trusted detection”.
- Ship through our MLOps platform so everything is reproducible, versioned, and auditable - not one-off scripts on your laptop. Someone else should be able to retrain and re-serve what you build.
- Work alongside our engineers and the analysts who understand the operational domain, and translate between the data and the mission - explaining to non-specialists what a technique can and cannot tell them.
Requirements
The Right Fit
- You are an applied data scientist, not a theory specialist. You have built and shipped detection on real datasets - not just toy or competition data, and you know the difference.
- You don’t reach for ML by reflex. You’re as comfortable with a deterministic rule, a statistical method, or a behaviour-based heuristic as with a model, and you choose the technique that actually fits the problem. The best engineers here invent the technique when none of the obvious ones work.
- You are fluent in Python and the scientific stack (NumPy, pandas, scikit-learn), and comfortable with at least one deep-learning framework (TensorFlow or PyTorch) when the problem warrants it.
- You have done anomaly detection, time-series, or geospatial work and have the scars to prove it - missing data, drift, imbalanced classes, and metrics that lie if you let them.
- You think about quality and drift, not just accuracy on day one. You know how to define what “good” means for a detection problem and how to keep it good as the data and the threat move underneath you.
- You are rigorous about evaluation. You can design a holdout that doesn’t leak, resist overfitting and metric-gaming, and you would rather report an honest 70% than a flattering-but-fake 95%.
- You write code others can build on. Version control, code review, reproducible pipelines - your work is not trapped in a notebook only you can run.
- You can own the pipeline from the source platform, not just the model. You have worked hands-on with the upstream geospatial and data-fusion platforms our data lives in and comes from, such as ArcGIS / ESRI, Palantir (Foundry or Gotham), and similar systems, and you can pull, shape, and trust data straight from them rather than waiting for someone to hand you a clean CSV.
- You are comfortable with ambiguity and partial data. “Find the vessels behaving like they have something to hide” is a brief you would find exciting, not paralysing.
- Bonus: Rust and edge compute. Some of our detection has to run at the edge, in constrained and disconnected environments. Familiarity with Rust, and with deploying to edge or embedded hardware, is a genuine plus.
- A quantitative background (physics, maths, engineering, computer science or similar) is an advantage but not required - demonstrated, shipped work counts for more than credentials.
Tech you will use
- Core: Python, NumPy, pandas, scikit-learn.
- Deep learning: TensorFlow or PyTorch (where the problem warrants it).
- Techniques: anomaly detection, time-series analysis, geospatial analysis, clustering, classification, and deterministic rules, statistical tests, and heuristics where those are the better tool.
- Production: Git, containerised and reproducible training, model serving (KServe), regression baselines, automated pipelines for fast and safe model/rule updates.
- Upstream platforms: geospatial and data-fusion systems such as ArcGIS / ESRI, Palantir (Foundry / Gotham), and similar - you’ll be pulling and shaping data straight from these.
- A bonus if you have it: Rust, and edge or embedded deployment.
- You’ll also encounter: maritime AIS data, signals/RF data, and deployment into secure and air-gapped environments.
You won’t need all of these on day one. We care more about how you think about a detection problem than which library you reach for.
What you will not be doing
- Building chatbots, LLM agents, or retrieval pipelines - that’s our AI Engineering team. This role is detection and modelling (machine-learning and deterministic) on structured sensor data. Your pipelines will enable fast, agentic delivery of detection, you won’t be building the agents themselves.
- Production backend service development.
- Building dashboards as an end in themselves.
Key Information
- We operate in defence and national security. Some of what we build remains classified and the work can be deeply sensitive.
- This is a small company (under 40 people). You will own the modelling layer of a product going into real operational use. Greenfield modelling on the maritime side, and hard open problems on the signals side. Nobody here models by committee.
- As OD scales, this role grows into senior and staff-level technical leadership of our data science work.
Benefits
Why Oxford Dynamics?
Join the most exciting growth area in the UK: AI and Robotics. Every member of the team has a major impact on the products and services we provide. Regardless of job title, you’ll get to make a real difference and learn from colleagues about all areas of our business.
Benefits
- Rapid career progression with meaningful ownership of core systems.
- Conference and tool budget.
- Flexible, Hybrid working model.
- Company pension (NEST, 4% employer contribution).
- Private healthcare.
- 29 days holiday plus public holidays.
Oxford Dynamics is committed to creating an inclusive team experience for all. Regardless of race, gender, religion, sexual orientation, age, disability, or parental status, we believe our work is at its best when everyone feels free to be their authentic self.
