vaniagarwal.dev
Switch to
Dark
Light
theme
Vani
Agarwal
i design, build, and evaluate AI systems
Now showing
01
02
03
04
Harness-Bench
on research.site
Benchmark, 4 harnesses, 12 questions · Nov 2026
DSE-246
on research.site
Benchmark, 2 search stacks, 246 tasks · Oct 2026
Astra-26
on research.site
Benchmark, 5 tool stacks, 26 tasks · Sep 2026
LBC-30
on research.site
Benchmark, 2 agents × 4 search APIs, 30 tasks · Aug 2026
Harness-Bench
tbd.chat
currently at
Past work
Designed a pay-as-you-go credit system and nudge engine for Copilot’s 10M+ student and consumer users.
Calibrated and shipped the production LLM-judge evals scoring agent replies across 2,000+ messages a day.
Engineered a document service unifying 3 data sources to serve 1,000+ PDFs a day.
Taught CSE 373 for 9 quarters, supporting 1,400+ students and leading teams of 12–15 TAs.
Writing
Why do benchmarks remain the industry’s public standard?
Essay, 16 Nov 2026
why am i embarrassed by my own creativity?
Essay, 13 Sep 2026
How do you restart your life in a city you already have history in?
Essay, 5 Sep 2026
Contact
X
LinkedIn
GitHub
v@internet.dev