Keras
Software Engineer III
- Leading the Keras model optimization charter.
- Designed and implemented the core Keras Quantization API, an extensible framework for cutting the memory footprint and compute requirements of large models.
- Implemented advanced low-bit quantization techniques (INT4, GPTQ, AWQ) in Keras and Qwix, reducing model size by 80% and VRAM usage by 50% while increasing token throughput by 48% with negligible accuracy loss.
- Designed and built an end-to-end benchmark framework for Qwix, an open-source model quantization library for JAX.
- Architected an end-to-end automated pipeline for KerasHub, removing all manual overhead behind 100K+ annual downloads for Google internal teams and saving hundreds of engineering hours.
- Designed and implemented migration paths from TensorFlow to JAX using Keras, improving adoption of next-generation ML frameworks across several teams.
Ads Data Layer
Software Engineer III
- Built the compiler, schema engine and planner components of a Relational Online Analytical Processing engine in C++.
- Re-architected the cube schema engine to support schema modularisation: schemas partitioned into discrete, independently managed modules with their own namespaces.
- Led work on a schema presubmit system enabling controlled launches of schema entities to production, for safe rollouts and deployment flexibility.
- Hosted interns and helped them succeed in their projects.
Software Engineer II
- Built core schema-engine features such as generated fields, user-defined functions and SQL module support.
- Built and maintained Java libraries and tools for accessing the cube schema, improving developer experience and productivity.