I build LLM agents and the data systems underneath them: pipelines, evals and analytics that show exactly what fails and what to fix.
Each dot is one run in production. The pink ones failed. This is roughly where the agent I took on at BrowserStack started.
Autonomous eval workflows on real production traces, so every failure is counted, not guessed.
Failures group into a few root causes. Fix the largest cluster first and gate every change on the eval suite.
Benchmark frontier models on accuracy, latency, cost and stability, then ship what the numbers say.
Reliable enough that customers stop noticing the agent and just use it.
Before agents, I built data platforms. It's why I treat AI reliability as a data problem: good evals are just good pipelines pointed at the right question.
Numbers from work that shipped, each one something I owned or led.
| Change | What moved | How | Where |
|---|---|---|---|
| 74% → 98% | Production AI agent stability | Autonomous eval workflows and failure analysis | BrowserStack |
| 40% → 88% | Ticket SLA adherence | Brought structure to a chaotic ops board | BrowserStack |
| +55% | User engagement | Shipped AI test-automation features end to end | BrowserStack |
| −40% | Deployment failures | Migrated multiple services into separate AWS EKS namespaces and set up their deployments | BrowserStack |
| −40% | Data processing cost | Airflow and Lambda ETL, 1M+ records a day | Refyne |
| −80% | ML training time | Training infrastructure on SageMaker | Refyne |
| 4h → 15m | Report generation | Automated reporting across 10+ sources | Refyne |
| +30% | Benchmark coverage | Knative support and two new serverless benchmarks | GSoC, SPCL |
Data, then infrastructure, then AI. Reliable agents need all three.
Feb 2025 to now
LLM agents for AI test automation, and the eval systems that make them dependable. Built the step-summarisation proof of concept and worked with Product from discovery to iteration.
Sep 2024 to Feb 2025
Airflow pipelines at 1M+ records a day, a batch ETL framework that handled 5× more volume, and ML training on SageMaker.
May to Sep 2024
Led the Knative integration in SeBS, the serverless benchmarking suite, and designed two new benchmarks.
May to Aug 2023
Built a URL caching system for KCIDB, KernelCI's database of Linux kernel test results, on GCP: Cloud Functions, Cloud Storage and Pub/Sub, a cache redirector, tests and documentation.
2020 to 2024
Mostly open source: GirlScript Summer of Code mentor, Microsoft Learn Student Ambassador, Linux Foundation LiFT Scholar.
Recognised for unusually fast growth and strong ownership: drove pivotal work on the AI step-summarisation proof of concept, helped push the AI agent toward 99% cloud-run stability, and resolved critical customer issues.
BrowserStack, Mar 2026Recognised for turning a chaotic ops board into a clear process that lifted ticket SLA adherence from about 40% to 88%, and for supporting the team's AI evaluation framework.
BrowserStack, Dec 2025Recommended by Marcin Copik for strong Python skills, picking up new cloud and serverless concepts quickly, and contributions that broadened what the SeBS benchmark suite can do.
SPCL, ETH Zürich, Sep 2024