Data Scientist
Role details
Job location
Tech stack
Job description
Assaia offers AI-based computer-vision solutions that monitor and analyse aircraft turnaround processes from video streams. Assaia's solutions help airports and airlines improve on-time performance, efficiency, safety, and sustainability metrics, helping them get their passengers to their destinations on time. We work with aviation companies across the world including New York JFKIAT, Toronto Pearson, United Airlines, British Airways, London Heathrow and more. See details and demos on our website and Linkedin!, We're looking for a Data Scientist to join Event Time Advancements, the team behind Assaia's flight event time and delay prediction capabilities - especially departure delay prediction and related operational forecasting problems. This is an applied ML role with end-to-end ownership. You'll work across the full lifecycle: building datasets and features from evolving event histories, training and evaluating models, and helping keep them reliable in live use together with backend engineering. A big part of the role is working on dynamic prediction problems rather than one-row-per-case modelling. The same flight or operational event evolves over time through repeated updates, and the model needs to make useful predictions from a changing picture as new information arrives. This makes the work especially interesting for someone who enjoys temporal modelling, sequential reasoning, and building systems that stay useful across multiple decision points. You should be comfortable making judgment calls on data trust, validation design, and trade-offs, and defending those calls to both technical and non-technical stakeholders. The team builds and owns its own tooling for dataset generation, training, and parts of real-time inference - so software engineering discipline (testing, review, deliberate design trade-offs) matters here as much as modelling skill. You'll directly shape live prediction quality for flight-event timing and delay-related capabilities used by operational teams, help determine which data is usable as ETA expands to new airports and prediction tasks, and turn one-off experiments into repeatable workflows the team can build on., + Train, tune, and evaluate models for departure delay prediction and related forecasts (event timing, duration, reason, occurrence).
-
Own tasks end-to-end: frame ambiguous operational questions as measurable ML problems, deliver evidence-backed recommendations.
-
Apply feature-based and neural approaches, including sequential models for evolving event histories. Data & Evaluation
-
Build datasets and features from flight status, turnarounds, weather, airport context, and operational detections.
-
Audit raw signals for reliability, timeliness, and usability.
-
Design leakage-free offline evaluation that reflects real product use and respects temporal structure. Production Ownership
-
Support inference code, monitor live quality, investigate regressions, and improve models when performance drifts.
-
Keep offline training aligned with live inference features and configuration; validate that offline results hold up in runtime.
-
Benchmark model output against existing estimates or baselines and explain trade-offs. Infrastructure & Collaboration
-
Build and maintain internal tooling for dataset generation, training pipelines, and inference infrastructure - treated as real infrastructure: tested, reviewed, and built with deliberate design trade-offs.
-
Collaborate with backend and other engineering teams on live-system integration.
-
Maintain reproducible project assets: configs, experiment code, notebooks, reports, monitoring outputs., Your input will be valuable and we actively ask everyone to share their thoughts and ideas to help steer the direction of the company's development. Additionally, we offer:
-
Participation in making important decisions, your ideas will be heard and implemented.
-
Always remote work and a flexible schedule.
-
Paid vacation, paid sick leaves.
-
Paid relevant courses/online education.
-
Great company culture based on honesty and mutual respect.
-
Live team events. We are a team of 120 now. We are looking forward to expanding it with you!
Requirements
-
Strong Python engineering skills - packaging (poetry/uv), testing, typing, and performance-aware code, with OOP applied judiciously where it earns its complexity rather than by default. Working SQL: comfortable with joins, aggregations, and filtering to extract data.
-
Writing typed, tested, maintainable Python that others can pick up, extend, and trust once it's running live - fluent in Git-based workflows, with good code-review judgment about when to be strict versus pragmatic.
-
Practical experience with the Python data stack: pandas or polars, numpy, pyarrow, tree-based models (CatBoost, XGBoost), and pydantic or attrs for data validation.
-
Able to use Docker and Kubernetes as a user, for running, debugging, and validating ML workloads.
-
Familiarity with agentic coding tools (e.g. Claude Code, Codex) and judgement about when to trust vs. verify their output. Modeling & Statistics
-
Strong judgement in feature engineering, validation design, metric selection, missingness handling, label reliability, and leakage prevention.
-
Solid grounding in statistics and probability, sufficient to reason about modelling choices and evaluation results.
-
Open to and capable of extending into sequential or neural approaches as the team grows this direction - practical understanding of training mechanics (losses, optimisation, schedulers, regularisation, stability) is valued even without deep prior neural-network experience (PyTorch experience is a plus).
-
Understands the pitfalls of time-based or event-based modelling - leakage, non-stationarity, evolving entities - through prior experience or strong first-principles reasoning. Ownership & Collaboration
-
Strong verbal and written English skills (B2+ level or higher).
-
Some direct exposure to models in production - you've shipped a model live, maintained one, or worked closely enough with serving and monitoring to understand what changes (and breaks) once a model moves from offline experimentation into a live system.
-
Comfortable owning a scoped, ambiguous problem end-to-end - structuring it, driving it to a useful conclusion, and knowing when to stop.
-
Clear communicator with both technical and non-technical stakeholders. Nice to have
-
Deep, hands-on fluency with horizon-dependent prediction problems across the full pipeline - framing the problem, constructing the dataset, training, and evaluating - going beyond a conceptual understanding of the pitfalls into practiced judgement at every stage.
-
Experience owning (not just using) retraining pipelines or drift-detection systems end-to-end.
-
Experience comparing offline and online model behaviour.