, world.
my name is harsh. i study cs and econ, and i like working on ml research, mostly around benchmarks and evals, model behavior, and reliability. lately, i've also been getting into mechanistic interpretability, experimenting with ways to understand what models are doing internally and what actually counts as a useful explanation of their behavior.
currently, i'm working with cambridge ai safety hub as a MARS V research fellow, studying whether llms preserve the intended meaning when translating natural-language statements into formal ones, and how we can detect cases where a formalization is technically valid but still unfaithful to the original statement.
here is some of what i've been working on:
research
- Deterministically Measuring Autoformalization Faithfulness measures whether llms preserve meaning when translating natural-language statements into formal ones, using deterministic grading instead of human or llm judges.neurips 26 · under review
- ToolFailBench is a diagnostic benchmark for understanding how llm agents fail at tool use, beyond just their overall performance.icml 26 · modal academic grant
- Emergent Misalignment in VLMs studies how narrow fine-tuning can lead to broader misalignment in vision-language models.iclr 26 · modal academic grant
- SafeGEPA asks whether optimizing prompts for performance quietly erodes alignment.in progress
things i've built
- LOFTE is an ios workout tracker with an ai coach, live on the app store.
- ScarTissue is a pr reviewer that remembers the bugs you already fixed.
- Drunk Buddy is an ai friend in your imessage that gets you home safe on a night out.
if i'm not working on either, you'll find me on a tennis court or at the gym.
✶✶✶✶
always down to talk new ideas and research. feel free to reach out below.