Geospatial AI/ML at AWS

Had the privilege of working 2020-2024 as an Data Scientist and Machine learning Engineer L6 at AWS in Public Sector, helping government and nonprofit with AI/ML workflows.

Had the pleasure of working with many talented peers, exceptional mentors and dedicated customers on an amazing diversity of use cases across verticals, disciplines, and technologies. I learned so much in such a short time (that yes also felt like eternity).

Overall, 3 main ‘eras’ in 4 years :

  • object classification, detection, tracking from video
  • object classification, detection, from satellite imagery (RGB, Radar / SAR)
  • Generative (LLM / RAG / Bedrock)

Other work included projects ranging from Sound Classification, Edge/IOT POC (Jetson & snowball), land mine removal support, disaster recover damage assessment, and even public AI/ML artwork

Had success with modules becoming re-usable artifacts across teams, as well as prospecting and scoping new POCs across verticals. Beyond my IC and tech lead roles, I identified and won projects that brought new work into our group.

Some work remains in part of this reference OSS solution for AI/ML processing of satellite imagery

source code : guidance for processing overhead imagery on aws and osml-models

so much of the work was customer data and I cannot share details.

However for very similar workflows with public documentation, see:

disaster response
crop segmentation
more custom models for disaster response

Tools: PyTorch, MXNet, SOTA model selection, training, data pre-processing, parallelization, model training and fine tuning, CUDA, runtime optimization, Arduino, Python, JupyterLab, SAMGeo

AWS Stack: CloudFormation/CDK, Lambda, SageMaker BYOC, S3, Redshift, Snowball edge, Bedrock, GovCloud


these public datasets were of great help for prototyping, pre-training, and evaluating computer vision models.

MOT Challenge PET and MOT17