Machine Learning Engineer

What you'll own

Scaling the run pipeline

• Turn existing model and ingestion code into orchestrated, parameterised pipelines that execute unattended across many sites, dates and configurations on GCP.
• GPU training and batch inference sized for whole-landscape runs. Throughput, cost per run, project size, and turnaround time are project KPIs.
• Experiment tracking and run provenance: every output traceable to its input data version, model version and config. When a number changes, we need to know why.
• The calibration loop where new field and/or LiDAR data arrive, triggering versioned retraining, with batch logs and regression checks against previous model versions.

Modelling and evaluation

• Design and run spatio-temporal ablation studies at scale. How much drone or field data is actually required to reach a given confidence level, and which input streams are carrying the prediction?
• Stress-test the models to find where they break: unseen biomes, persistent cloud cover, sparse ground truth, unusual canopy structures. How much data do we need and how can we turn this into a sampling plan for our customers?
• Implement uncertainty quantification: per-pixel confidence intervals, and propagation of input and model uncertainty through to the reported carbon figure. Evaluation on calibration and probabilistic scoring.
• Interrogate architecture choices : one model across biomes, or specialists per ecosystem?
• Work with learned representations from geospatial foundation models, and help us evaluate candidate alternatives on evidence rather than on hype.
• Build the multi-modal fusion path: co-registering Sentinel-2, VHR optical (Planet/Maxar), Sentinel-1 SAR, GEDI/ICESat-2 spaceborne LiDAR, drone LiDAR and field inventory into coherent training and evaluation sets.

Requirements

Essential:

• Strong Python. Comfortable with the geospatial stack (xarray, rasterio/GDAL, geopandas, dask or equivalent) and a modern deep learning framework (currently, we're in pytorch/jax but are flexible).
• Demonstrable experience training and evaluating ML models on spatial or other large gridded data
• Track record of taking research code to reliable automated execution: orchestration, containerisation, config management, CI, cloud batch/GPU compute. This is the single most important requirement.
• Experience with workflow orchestration and experiment tracking tooling (Airflow/Dagster/Prefect, Vertex AI Pipelines, MLflow/W&B, or equivalents).
• Good grasp of uncertainty: you understand different uncertainty propagation and quantification principles and you care which one a stakeholder is asking for.
• Rigorous validation instincts. You default to spatially and temporally held-out evaluation because you know random splits leak in geospatial data.
• Software engineering fundamentals: version control, testing, code review, reproducible environments.

Desirable:

• GCP specifically (Vertex AI, Batch, GKE, BigQuery, Cloud Storage) and infrastructure-as-code. This is where we work currently.
• Experience working with pretrained representations or embeddings from foundation models, including their failure modes.
• Familiarity with LiDAR processing (multi-return point clouds, canopy height models) and forest inventory data (DBH, allometry).
• Exposure to carbon MRV or the relevant standards (Verra, Gold Standard, Isometric, IPCC LULUCF guidance).
• Published or otherwise peer-reviewed technical work in remote sensing, ecology or ML.

Characteristics we like

• Treats automation as a research tool
• Quantifies doubt rather than hiding it. Our commercial value is a credible error bar.
• Empirical over enthusiastic. Comfortable stopping an approach that doesn't beat the benchmark.
• Owns the whole path. Equally at home in a notebook, a Dockerfile, a Terraform config and a field-data QA review.
• Writes things down. Model cards, methodology write-ups and QA documentation are deliverables here as our customers need a paper trail.
• Works well with messy ground truth. Field data arrives late, inconsistent, and in the wrong CRS.
• Comfortable in a small team. Five+ technical people.

Success in the first 12 months

• Model runs are automatic.
• Every run is reproducible and traceable to its data and model versions.
• Ablation evidence exists that tells a project developer how much field data they need to buy for a given confidence level.
• Confidence/credible intervals generated programmatically for every biomass estimate, in a form external verifiers accept.
• We know, with evidence, the biggest weaknesses in the current model

🌍 Location & hours

UK-based, with core hours 10:00–16:00 for collaboration. Fully remote for candidates outside commuting distance of London (with a London visit approximately once every 8 weeks); if you're based near London, you’re welcome to co-work with us in central London 2 days a week.

📄 Contract

• Engagement: Full-time (ML Engineer)
• Time commitment: 5 days/week
• Start date: ASAP
• Gross Salary: Competitive (paid in GBP)
• PTO: 25 + Bank holidays
• Equipment: A company laptop can be provided if needed

🎉 Ready to Apply?

Excited by the chance to build software that measurably helps nature projects get off the ground? We’d love to hear from you — even if your profile isn’t a perfect match.