Synthetic Aperture Radar (SAR)
Speckle-resilient backbones and custom attention mechanisms that interpret polarimetric backscatter signatures.
Applications
Computer vision · Geospatial AI · Aerial
Deep tech. Planetary insights.
We are a deep-tech research startup working at the intersection of advanced computer vision, geospatial AI and aerial data systems. We don't just apply existing models. We push how neural networks interpret the Earth, turning raw feeds from space, aerial and land into high-fidelity, actionable intelligence.
SpaceSatellite SAR, hyperspectral and optical
AerialAerial imagery and LiDAR point clouds
LandGround-level sensor feeds
Land-cover segmentation
A model labelling water, cropland and built-up areas
Multispectral · Tile 14 / 208
Water · 0.96
Cropland · 0.91
Urban · 0.93
All-weatherRadar that sees through cloud
Sub-pixelDetail from hundreds of bands
Data we work with
SAR
Hyperspectral
Multispectral
Optical
Aerial LiDAR
Core research pillars
Built for the three major modalities of remote sensing
We develop next-generation segmentation, classification and object detection frameworks, each optimised for how its sensor sees the world.
Hyperspectral cube analysis
Bespoke 3D-CNNs and spectral-spatial transformers that read spatial features and continuous chemical curves together, without overfitting.
Applications
Multispectral & optical delineation
Zero-shot foundation models and cross-attention architectures that align different band resolutions seamlessly.
Applications
The algorithms we are advancing
State-of-the-art vision, re-engineered for the Earth
We adapt leading architectures so they work on the data remote sensing actually produces, from radar phase grids to raw 3D point clouds.
- 01
SegFormer extensions
Hierarchical transformer encoders that model long-range context, cutting the compute needed for ultra-high-resolution multispectral grids.
- 02
Geospatial SAM adapters
Custom adapter layers for foundation models like SAM 2, bridging everyday vision and non-optical data such as SAR phase grids.
- 03
PointNet++ for aerial LiDAR
Point-set networks that ingest unordered 3D point clouds directly, for canopy profiling and digital terrain models.
- 04
Multi-scale DeepLabv3+
Tuned atrous spatial pyramid pooling that captures winding rivers and pipelines alongside macro-scale features, without edge erosion.
Partner with Gradient Consulting
The research infrastructure your data needs
From a single feasibility question to a production-ready model, we work alongside your team and your sensors.
Custom AI model design
Neural networks built for your payload
Purpose-built models tailored to your specific sensor and the questions you need it to answer.
Data pipeline optimisation
Pre-processing that scales
Scalable pre-processing, pansharpening and noise reduction engines for large imagery archives.
Feasibility & R&D audits
Research, tested on real data
We validate state-of-the-art vision papers against real-world geospatial datasets before you invest.
Who we work with
Teams with complex sensor data and hard questions
- Aerial enterprises unlocking the potential of a proprietary constellation
- Environmental agencies tracking critical climate change
- Ventures seeking deep-tech validation for a product or thesis
Have a sensor feed worth decoding?
Tell us about your data and the question you need answered. We'll tell you honestly what is feasible, and how we would approach it.