Apr 2022 – Nov 2025 · 3 yrs 7 mos

• Observability and performance-focused engineering of Theseus, Voltron Data’s flagship accelerator-native distributed query engine, providing and acting on semantically rich insights.
• Propose, help design and implement a priority-based composable node-level task scheduler for Theseus, to reduce data movement and optimize GPU memory utilization, by gaining explicit and tunable control over execution, providing immediate improvements to stability and reduction in technical debt while enabling and contributing new improvements, such as concurrent query execution.
• Introduce observability into Theseus, by helping instrument the code base with OpenTelemetry, establishing a semantic convention and architecting the post-processing and analysis pipeline to gain insights into Theseus.
• Overcome scale and engine behavior modeling constraints imposed by OpenTelemetry by designing, implementing and integrating a tailor-made finite-state machine based in-house specification enabling always-on low-overhead distributed profiling and live query progress, making idiomatic use of Rust and Apache Arrow, from instrumentation to post-processing, providing fine-grained and actionable insights.
• Perform extensive experiments to identify bottlenecks and alleviate them, by optimizing various control systems.

Jan 2022 – Apr 2022 · 3 mos

• Worked on Apache Arrow, the columnar format for fast data interchange and in-memory analytics, in form of implementing compute kernels.
• Benchmarked the performance of Apache Arrow compute kernels on various on-prem and cloud architectures, investigating the correlation between performance and architectural differences.
• Explored the use of SIMD to accelerate compute heavy portions of compute kernels.
• Helped refactor the logging subsystem of Theseus, Voltron Data's distributed accelerator native query engine, to use type-safe APIs, resulting in well specified logs that can be programmatically analyzed to obtain better insights.

May 2021 – Aug 2021 · 3 mos · Remote

• Investigate, implement and integrate a superior spectral estimation algorithm called the Multitaper Periodogram into a scientific software package named Stingray.
• Extend this algorithm to support time series with uneven temporal sampling, implementing a recent research paper which builds on the popular Lomb‑Scargle periodogram.
• Optimize the Lomb‑Scargle periodogram’s implementation to run in O(n log n) time complexity as opposed to O(n^2), following a research paper by Press and Rybicki, making it vastly more relevant for practical research.
• Test and benchmark all the implemented algorithms.