Chandra Suvarna

Google Research

• Developed pretrained embeddings for geospatial regions to capture semantic and structural properties of the built environment
• Designed and constructed a large-scale geospatial dataset from scratch, including feature engineering, labeling strategies, and data validation
• Implemented and trained embedding models in TensorFlow, iterating on architectures and training objectives
• Co-author on a peer-reviewed publication resulting from this work (ACM)

Gemini for Geospatial Applications

• Designed and established comprehensive benchmarking frameworks to systematically measure Gemini’s performance on geospatial reasoning tasks
• Defined and implemented evaluation metrics to quantify model quality, identify failure modes, and track performance improvements over time
• Built and curated high‑quality training datasets to support post‑training and fine‑tuning of Gemini for geospatial use cases
• Collaborated closely with Google Research to analyze evaluation results, iterate on data and modeling strategies, and drive measurable improvements in Gemini performance

Google Maps (ML Systems & Infrastructure)

• Worked with SOTA ML models such as: Vision transformers, LLMs, object‬ detection models, embedding generation models‬ to geolocate objects of interest for Google Maps
• Built and maintained infrastructure for ML dataset storage, exploration, and lifecycle management
• Supported large‑scale ML pipelines with an emphasis on data quality, reproducibility, and reliability

• Worked on building a composable system for performing data validation
• Built a core library into existing API using Spring and Java that reads in a spec containing data validations, and runs all validations listed against data provided

• Built a Spring Boot web service that automates the process of setting up monitoring and alerting in APMs (Datadog, New Relic) for deployment ready services
• Used New Relic and Datadog APIs to set up dashboards, graphs of metric, and alerts for service
• Integrated usage of telemetry for a web service (throughput, average response time) to monitor service
• Utilized Swagger Documentation to collect the service's API and endpoint information