About the role
RoleYou will be the architect of the model’s latent space, designing foundation models for multi-spectral, multi-temporal, and multi-resolution geospatial data.This is a hands-on role involving prototyping, experimentation, and large-scale training. You will work across representation learning, model scaling, and spatiotemporal modeling to build systems that generalize across sensors, geographies, and time.
ResponsibilitiesRepresentation Learning
- Design and implement self-supervised learning (SSL) objectives (e.g., Masked Autoencoders, DINO-style methods, contrastive learning) tailored for geospatial data
- Develop multi-modal representations spanning optical, SAR, elevation, and derived signals
- Ensure representations transfer effectively across tasks such as segmentation, classification, and change detection
- Design evaluation strategies to measure generalization across geographies, sensors, and time
- Design and scale models based on Vision Transformers (ViT), hybrid architectures, or State Space Models (e.g., Mamba) to large parameter regimes
- Apply modern training techniques such as RMSNorm, FlashAttention, mixed precision, and gradient checkpointing
- Run scaling experiments, ablations, and architecture explorations grounded in empirical rigor
- Leverage insights from scaling behavior to make compute-efficient decisions across model size, data, and training strategy
- Develop methods to model time-series satellite data, capturing:
- Seasonal patterns
- Temporal dependencies
- Long-term land-use changes
- Explore sequence modeling, memory mechanisms, and temporal tokenization strategies
- Design ML systems as end-to-end pipelines (data ingestion → curation → training → evaluation → deployment → feedback)
- Make explicit trade-offs between model quality, latency, cost, and data freshness
- Work with platform teams to optimize:
- Distributed training (FSDP, DeepSpeed)
- GPU utilization
- Data pipelines and experiment throughput
- Build reusable components and abstractions, not one-off models
Preferred BackgroundExperience
- 5–8 years of experience in ML research or applied research roles
- Experience in large-scale foundation model development (vision, multimodal, speech, or related domains)
- Experience training and/or fine-tuning billion-parameter models
- Experience working with sequence, video, or temporal data
- Exposure to geospatial foundation models such as:
- Prithvi
- Clay
- Segment Anything Model (SAM) (nice to have)
- Expert-level proficiency in PyTorch or JAX
- Strong experience with:
- Distributed training (FSDP / DeepSpeed)
- Large-scale datasets and training pipelines
- Familiarity with transformer architectures and training dynamics
- Bonus: CUDA / performance optimization experience
- Familiarity with efficient scaling techniques (e.g., Mixture of Experts) is a plus
- Strong experimental rigor and ability to design meaningful ablations
- Track record of publishing or contributing to state-of-the-art research in representation learning or generative modeling