At Marvedge, I integrated TalkNet into a video-analysis pipeline that processed video at 28.5× real time on a T4 GPU. I also accelerated preprocessing 53× by replacing random-seek decoding with a sequential FFmpeg pipeline.
I cut production cloud costs from $1,700+ to about $60–80 per month by moving video delivery to Cloudflare R2 and replacing always-on GPU instances with Spot Auto Scaling Groups. I also built multilingual speech and LLM pipelines, including subtitle translation checks for 30+ target languages.
In the LLVM Project, I contributed a merged MLIR bufferization fix and have three open Clang Static Analyzer PRs. My FFmpeg contributions include two open AArch64 NEON PRs with passing CI and measured 14–30× speedups; I also authored a preprint on verifying AI-generated code.
