AI/ML Engineer – Earth Observation

Who We Are Kimedes AI is a climate-tech deeptech company founded in December 2024 in Arenys de Munt, on the Maresme coast of Catalonia, Spain. Europe loses around 25% of its treated drinking water before it reaches the tap—leaked through buried pipes in decades-old networks. Excavating to find these leaks is slow, expensive, and largely guesswork. We took a different path: instead of deploying hardware underground, we built a data cube over the network—combining satellite and environmental data from multiple independent sources, stacked over time—read by our AI engine to detect the subtle signature of an underground leak. No single data source suffices; making them agree is the core challenge. From there, we’re building tools to act on these insights. Our flagship platform, Syracusa, turns leak detections into inspection plans and regulatory audit workflows that municipal operators can execute the same week. We’re four founders with paying production customers, recognized by the European Commission with GovTech4All Silver 2026, supported by the European Space Agency (ESA), selected in ACCIÓ’s Top 25, and granted access to supercomputing at the Barcelona Supercomputing Center. Scientific collaboration is led with CTTC and IEEC. The Role, Straightforward Think. This role is about thinking. We have a problem—and we need you to think deeply about it. We believe the solution lies in a deep learning model, and we want you to lead it: own the model and the development of supporting tools. Not maintain them—but build them: design, productionize, and define maintenance workflows. This is the exciting part: helping shape Kimedes, not just keep it running. We have an R&D project funded by the ESA with contractual deliverables and deadlines, and a follow-up national 3-year project. The AI component of both is yours. You’ll start from a blank slate, with access to supercomputing, remote sensing physicists at your disposal, and a first deliverable in five months. Your architectural decisions will become documents reviewed by the ESA; your model’s predictions will trigger municipal crews breaking open streets. Both carry real consequences. We’re a small team, so we offer high autonomy and expect strong problem-solving skills. That’s exactly what we provide: space to grow professionally, freedom to implement your ideas. What You’ll Do Design Define the architecture of a model that ingests multi-source, multi-temporal satellite data over the same territory and outputs leak probability—spatial and temporal fusion, handling irregular acquisition schedules. Contribute to AI design decisions, grounding them in physical reality: adapting loss functions or transformer-based architectures, tuning variables, or designing data preprocessing to improve model metrics. Help design the broader ecosystem we’re building around the core model. Document architectural decisions—including discarded alternatives and rationale. This isn’t bureaucracy: it’s what external technical reviewers assess. Train Build the training pipeline and architecture tournament, with a written promotion rule: a candidate configuration enters only if it significantly beats the classical baseline and* preserves calibration. Tackle severe class imbalance: leak events are rare, and models tend to learn "no leak." Calibrate probabilities and report metrics adjusted to natural prevalence. Scale training at the Barcelona Supercomputing Center using distributed training and ablation studies to quantify each data source’s contribution. Deploy to Production Package the inference engine as a service with an API, so the platform can consume it and a municipal technician receives a usable map the same day. Support field validation: compare model predictions with field findings, understand failure modes, and recalibrate. This is what separates a pretty model from one that works. What We’re Looking For Requirements At least 4 years of experience training models used in production (notebooks don’t count). Strong Python and PyTorch skills. Experience designing architectures—not just using them—you should be able to defend your choices and explain what you discarded. Experience with spatio-temporal or multimodal data, and understanding why random splits are misleading when spatial correlation exists. Deployed models in production—someone has relied on your predictions. Working English: deliverables go to the ESA, scientific meetings are technical and conducted entirely in English. Highly Valued Remote sensing: radar, interferometry, coherence, optical multispectral. If you come from this field, this role was written for you. GIS knowledge (PostGIS or similar). HPC or distributed training experience. Handling severe class imbalance and probability calibration. Applied physics or modeling background—for physical consistency. PhD. Not required, but especially well-suited for this role. Not Required Water sector experience—you’ll learn it in the first pilot. Team management experience—this is a hands-on role, not a leadership one. What We Value Most We’re a small team that has gone far through hard work. We seek proactive people who propose solutions to problems, aren’t shy to ask basic questions in meetings, and can’t sleep until they understand why a result came out a certain way. Curiosity about the physical reality of what we measure, and enthusiasm for testing unconventional ideas. What We Offer Your Time Is Yours 38.5-hour workweek (Spanish legal standard: 40 hours). We chose less—and wrote it into the contract. Friday: 6.5 hours, uninterrupted. All of August: 6.5 hours daily. 25 days of vacation + your birthday off. Flexible start time between 8:00 and 10:00. ~1,686 productive hours/year vs. 1,800 in the collective agreement. Hybrid, With Clear Rules 2 days in our office in Arenys de Munt, 3 days remote from your chosen location. Up to 4 weeks/year may require full on-site presence—only for objective reasons and with written notice. Four weeks is the ceiling. Full home and office workstation provided and paid for by us. Written remote work agreement from day one, with monthly compensation for internet and electricity. No Messages After 6:00 PM Your end time follows your start time, but our communications don’t. If something arises at 7:30 PM, it’s scheduled for next morning. Not “you’re not obliged to reply”—it’s simply not sent. Tools & Science MacBook and full workstation delivered on day one. Claude Pro for everyone: AI is part of how we work, not a rationed extra. Direct access to supercomputing—no shared GPUs. You will publish. We collaborate with CTTC, IEEC, and the Barcelona Supercomputing Center—and research that stays in a drawer is wasted research. Conferences are evaluated case-by-case: if it matters to our mission, you go and present. How We Work Together Regular 1:1s with founders (who sit right next to you). Real voice in company direction—strategy doesn’t come from a higher floor at our size. Monthly demo days, one offsite/year in Catalonia, team lunches on the first Friday of each quarter, and a year-end dinner. What we won’t promise: foosball, offsites in Marrakech, or fluffy mission statements about passion and excellence. We’re a handful of people with satellites, real customers, and a serious problem. Salary Range €38,000 gross annually Selection Process Four stages, completed in three weeks. Initial call, 30 minutes. We present the project; you tell us what you’re looking for. Design discussion, 90 minutes. We present the real problem on the table and think through it together on a whiteboard. No right answer—and no coding. We want to see how you reason when the problem is unsolved. No 20-hour unpaid take-home tests. Scientific conversation, 60 minutes, with our research partners. You ask them too. Final conversation with founders, 45 minutes, followed by an offer. We reply to everyone—whether there’s a fit or not. Where We Are Arenys de Munt, Maresme—40 minutes along the coast from Barcelona. Mountains behind, sea in front, and a commute that doesn’t eat your life. How to Apply Write to us telling us about a model you built that you’re proud of—and what you’d do differently today. That tells us more than any cover letter. At Kimedes, we select based on what you can do and how you think—not your gender, age, origin, orientation, or any other trait unrelated to the job. If you need accommodations during the process, let us know and we’ll adapt. Data Controller: Kimedes Artificial Intelligence S.L., B75632604. Purpose: Manage this recruitment process. We retain your data for 2 months after the process closes. You may exercise your rights of access, rectification, deletion, opposition, limitation, and data portability at [email protected]. More info: [link to privacy policy]