Demand Estimation

Rollout on gigabytes of image

(c) Headline image from ThinkingMachine blogpost

Project context: External blogpost

  • A comparable project as above but with larger regional provider, anonymized to preserve client identity.

Personal contributions:

  • Built, trained, and deployed ML models rolled out to 50+ GB worth of satellite images to generate previously unknowable and infeasible to collect data, saving $15/sqkm vs commercial API for regional telco client;
  • Surpassed critical usability threshold for an AI model for a international telecommunications client, by enhancing accuracy by 13% through error analysis and novel graph-based solution;
  • Achieved acceptance criteria for enterprise ML CV project, allowing on-time project delivery, by successfully negotiating image-based error-correction web-app feature with client;
  • Rolled out ML model on 50+ GB worth of satellite images in 2 hours by distributing compute on 40 VMs via GCP DataFlow, vs usual 2 days on 1 machine to ensure on-time delivery;
  • Reduced rollout runtime by 2mins, by optimizing graph calculation with cached hashmap;
  • Productionalized SQL pipeline with automated data quality checks via DataForm;
  • Automated static analysis, profiling and test coverage for ML APIs;
  • Prototyped building classification model, including modifying PyTorch layers to enable fusion of image and tabular features to one model;
  • Developed and productionalized object detection data product, saving $15/sqkm for client rollout instead of using a commercial API;
  • Developed and productionalized building height prediction model, saving $59/sqkm for client rollout;
  • Implemented a graph-based model feature for building use classification model, used in final rollout version;
  • Taught fundamental ML/DL to external clients, contributing good relationship for return engagement;