I build the parts of the backend that decide where requests go.

Worked on the service mesh that routes traffic between Stripe's internal services. The mesh picked backends in a way that couldn't tell when one was overloaded, so I rebuilt it around power-of-two-choices weighted by the CPU and queue depth each backend reports. Overload errors dropped 40% across 50+ services handling 400M+ requests a day, and the services could stop over-provisioning for their hottest replica, which came to roughly $200K a year.

Then added zone-aware routing that spills traffic to neighboring zones when one gets hot, cutting p99 latency 23%, and ran the production rollout through feature-flagged canary ramps gated on error and latency metrics. No incidents.

Python data pipelines for internal workflows: a scheduled ETL for payroll reconciliation with schema validation and fuzzy record matching across Excel and shared-drive sources, and a network log pipeline that sessionizes raw logs into SQL for Power BI dashboards.

Benchmarked six fine-tuned BERT variants with a reusable Hugging Face evaluation script, exported the best to ONNX with int8 quantization for 3× faster CPU inference, and wrote the data checks that removed 5% noisy labels from a 1M+ ticket dataset, raising baseline F1 by 2 points.