Sai
M A Sai Adithyaa
B.Tech CS student at Amrita Vishwa Vidyapeetham, Chennai, batch of 2027. I build machine-learning tooling, write about things I had to figure out the hard way, and solve problems on Codeforces when I should be sleeping.
current research / work focus
I'm building an end-to-end ML pipeline on tribology data for PA6/PA66 polyamide composites. The bottleneck in this field isn't modeling — it's data. Wear-rate measurements are scattered across decades of papers, each with its own notation, tables, and figures. I'm extracting structured records from roughly 35 papers using a Gemini 1.5 Flash vision pipeline, then validating everything by hand before a single model gets fit.
On the modeling side the plan is random forests, XGBoost, and Gaussian process regression, with GroupKFold cross-validation grouped by source paper. The grouping matters: samples from the same paper share equipment and protocol, so a random split quietly leaks and inflates your scores. The point of the project is honest generalization estimates on genuinely small, noisy data — not leaderboard numbers.
A random split on multi-paper data measures nothing but your test-set luck. — [attribution / source]
# pipeline, roughly papers (n≈35) --vision--> gemini-1.5-flash --extract--> structured records records --manual validation--> dataset dataset --models--> RF · XGBoost · GPR cv: GroupKFold(k, groups=paper_id)
open source
- Spectrum UI: CLI — shipped improvements to the component library's CLI tooling, merged as PRs #74 and #75.
- InsForge — merged PR #502 into the open-source platform for isolated API, integration, and E2E test environments.
- Goose — merged PR #5381, contributing to the AI agent developer-tool ecosystem.
- Chimoney Community Projects — merged PR #532 in the Chimoney open-source community repository.