Experience
Where I’ve built
I've had the chance to work on interesting projects in the industry and academia. Click on a role to see more details!
Industry
- Built a memory sanitizer/profiler from scratch for Microsoft’s MAIA Triton accelerator compiler (MLIR with custom dialects); used with a custom harness to surface 4+ bugs/optimizations.
- Optimized kernels for Google’s Gemma 2 model to run ~2× faster on the MAIA chip versus PyTorch compilation.
- Coordinated complex changes across a hierarchy of 8+ systems for a project to double Google’s TPU delivery capacity.
- Designed a system to bill GCP customers in protected locations with heavy data-transfer restrictions.
- Exceeded core scope to earn 16 peer/leadership nominated awards for unblocking critical issues and accelerating cross-team collaboration.
- Built and maintained big-data pipelines processing over 60 billion records/hour for resource planning.
- Reduced pipeline computation time from 14h to 2h (~86%) by optimally partitioning shards to reduce stragglers.
- Trained language models with domain adaptation and auxiliary tasks for NER in long-form documents.
Academic Research
- Fine-tuned and benchmarked vision models (VLM, VideoMAE) for interpretable, few-shot classification.
- Built an end-to-end production-ready pipeline for data processing and model inference on 10k+ videos.
- Advisor: Dr. Deva Ramanan
- Developed a framework using foundation image generative models to guide targeted, sample-efficient data collection — up to 13% AP50 improvement while reducing required training data.
- Advisor: Dr. Fernando de la Torre
- Developed obstacle-avoidance algorithms using optical-flow-based visual servoing. Advisor: Dr. K. Madhava Krishna
- Improved multi-view pedestrian detection/tracking with visual transformers. Advisor: Dr. Anoop Namboodiri
Microsoft
Redmond, WA
