Research Scientist — 3D Reconstruction & Spatial AI

We are a well capitalized stealth VC-backed startup building a new type of spatial AI capable of universally solving autonomy. We innovate at the foundational layer of AI by training our own AI models.

Our team is composed of AI pioneers and leaders from Google X, Google Brain, and Space Agencies. Several of us are repeat founders, with deep commercialization insights across multiple enterprise segments. We enjoy long, deep ideations around entirely unexplored AI use cases in autonomy (cars, drones, robots).

• Curious: you have an innate curiosity for all things intellectual, and it’s something you can’t turn off. You obsess over finer details others miss.
• High intensity: you thrive in a high-stakes environment, and are driven by an innate obsession, not by others.
• Fast learner: you gravitate toward learning new things, and often find yourself learning more quickly than everyone around you.
• Driven by discomfort: you enjoy leaving your comfort zone and challenging yourself
• Creative: track record of solving hard problems with solutions worthy of academic papers
• Educator: you take pride in your ability to communicate complex topics clearly and have excellent speaking and writing skills.
• Zero ego: you don’t just take feedback, but truly see it as a gift. You don’t wait until feedback is given, but solicit it with every opportunity.
• Founder mentality: you roll up your sleeves to help solve the most pressing problem on a given day, even if it has nothing to do with this job post.

San Francisco, CA

• 3D Reconstruction: Real depth in multi-view geometry and modern learned reconstruction. Feed-forward pointmap models, neural or Gaussian scene representations, monocular and sparse-view depth, and a clear view of where each one stops working.
• Publication Record: First-author work at CVPR, ICCV, ECCV, NeurIPS, ICLR, CoRL, RSS or equivalent. We care that other researchers have had to reckon with your results.
• Metric Scale: Getting real-world scale and georeferencing out of monocular or sparse input, not up-to-scale reconstructions that need a post-hoc fix.
• Uncertainty: Calibrated confidence, so the system reports that its geometry is wrong before something downstream acts on it.
• Degraded Conditions: Low texture, specular and moving surfaces, glare, repetitive structure, thermal and other non-visible modalities.
• Deployment: You have taken a model from a paper to a vehicle. Jetson-class embedded targets, quantisation, latency budgets. You do not consider this someone else's job.
• Programming: Strong Python and PyTorch, comfortable in C++.
• Domain-Specific: Aerial or maritime reconstruction, photogrammetry, remote sensing, SfM at scale, or SLAM.