Location : Bangalore, India

Employment Type : Full-Time

About SatSure :

SatSure is a deep-tech Earth intelligence and decision intelligence company operating at the nexus of agriculture, infrastructure, and climate action. We turn Earth Observation data into actionable insights for governments, financial institutions, enterprises, and millions of end users, with a strong focus on developing markets.

As part of this mission, we are building geospatial foundation models that learn directly from Earth Observation data - including optical, SAR, elevation, and multi-temporal satellite imagery - at scale.

This role sits at the heart of that effort. You will architect and train large-scale models capable of generalizing across geographies, sensors, resolutions, and time, helping shape the core intelligence layer behind SatSure's geospatial products rather than simply fine-tuning existing models.

Role :

You 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 deeply hands-on research and engineering role involving prototyping, experimentation, architecture exploration, and large-scale model training.

You will work across representation learning, model scaling, multimodal learning, and spatiotemporal modeling to build reusable ML systems that generalize across sensors, geographies, seasons, and time.

Key Responsibilities :

Representation Learning :

- Design and implement self-supervised learning (SSL) objectives tailored for geospatial data, including approaches such as Masked Autoencoders, DINO-style methods, and contrastive learning.

- Develop rich multi-modal representations spanning optical imagery, SAR, elevation, and other derived Earth Observation signals.

- Build representations that transfer effectively across downstream tasks including segmentation, classification, change detection, and related geospatial applications.

- Design robust evaluation strategies to measure generalization across geographies, sensors, resolutions, seasons, and time.

Model Development & Scaling :

- Design, build, and scale foundation models based on Vision Transformers (ViT), hybrid architectures, and State Space Models such as Mamba to large parameter regimes.

- Apply modern model-training and optimization techniques including RMSNorm, FlashAttention, mixed precision, and gradient checkpointing.

- Conduct rigorous scaling experiments, architectural explorations, ablations, and benchmarking to understand model behavior and performance.

- Use insights from scaling behavior to make compute-efficient decisions across model size, dataset composition, architecture, and training strategy.

- Explore efficient scaling techniques such as Mixture of Experts (MoE) where applicable.

Temporal & Spatiotemporal Modeling :

- Develop methods for modeling time-series satellite and Earth Observation data, capturing :

1. Seasonal patterns

2. Temporal dependencies

3. Long-term land-use changes

4. Changes across observations and sensors

- Explore advanced approaches to sequence modeling, memory mechanisms, temporal representations, and temporal tokenization.

- Develop models capable of learning meaningful representations from complex multi-temporal and multi-sensor datasets.

Systems-Level Thinking :

- Design ML systems as complete end-to-end pipelines, covering : Data ingestion - Data curation - Training - Evaluation - Deployment - Feedback

- Build systems and research workflows that can move beyond experiments into reliable, reusable ML capabilities.

- Make explicit engineering and research trade-offs between model quality, latency, computational cost, data freshness, and scalability.

- Work closely with platform and engineering teams to optimize :

1. Distributed training

2. FSDP / DeepSpeed

3. GPU utilization

4. Large-scale data pipelines

5. Training throughput

6. Experiment throughput

- Build reusable components, frameworks, and abstractions rather than one-off models or experiments.

- Contribute to the development of scalable training infrastructure capable of supporting increasingly large geospatial foundation models.

Preferred Background :

Experience :

- 3 - 8 years of experience in ML research, applied research, or closely related roles.

- Demonstrated experience in large-scale foundation model development, whether in vision, multimodal learning, speech, language, video, or related domains.

- Experience training and/or fine-tuning billion-parameter models.

- Strong experience working with sequence, video, temporal, or spatiotemporal data.

- Experience taking research ideas from hypothesis and experimentation through meaningful evaluation and scalable implementation.

- Exposure to geospatial foundation models or related architectures such as :

1. Prithvi

2. Clay

3. Segment Anything Model (SAM)

Geospatial experience is advantageous but candidates with strong foundation-model research experience in adjacent domains are also relevant.

Technical Skills :

- Expert-level proficiency in PyTorch or JAX.

- Strong understanding of Transformer architectures, representation learning, and modern model-training dynamics.

- Strong experience with :

1. Distributed training

2. FSDP

3. DeepSpeed

4. Large-scale datasets

5. Large-scale model-training pipelines

6. Multi-GPU training

- Ability to reason about model architecture, optimization, training stability, computational efficiency, and scaling behavior.

- Experience designing rigorous experiments and interpreting results beyond headline model metrics.

- CUDA and low-level performance optimization experience is a strong advantage.

Additional Strengths :

- Familiarity with efficient model-scaling approaches such as Mixture of Experts (MoE).

- Strong experimental rigor with the ability to design meaningful ablations, baselines, evaluation protocols, and scaling experiments.

- Strong understanding of self-supervised learning, representation learning, multimodal learning, and generative modeling.

- Ability to independently explore new architectures and research directions while translating promising ideas into practical systems.

- Track record of publishing, implementing, or contributing to state-of-the-art research in areas such as representation learning, foundation models, multimodal learning, generative modeling, computer vision, or related domains.

- Research publications or meaningful contributions to the broader ML research ecosystem are highly valued.

Impact of the Role :

This role will help shape the core intelligence layer of SatSure's Earth Observation platform.

You will work on foundation models that learn from massive volumes of multi-sensor, multi-temporal geospatial data and are designed to generalize across regions, seasons, sensors, and applications.

The resulting models will power downstream intelligence across areas including agriculture, infrastructure, climate action, environmental monitoring, and other Earth Observation applications, enabling SatSure to deliver reliable geospatial insights at scale.

Benefits :

- Medical Health Cover for you and your family, including unlimited online doctor consultations.

- Access to mental health experts for you and your family.

- Dedicated allowances for learning and skill development.

- Comprehensive leave policy including casual leaves, paid leaves, marriage leaves, and bereavement leaves.