Amirali A.

Key Skills

Software

No software listed

Top Subject Matter

Physics

Mathematics

Top Data Types

Image

Text

Video

Top Task Types

Fine-tuning

RLHF

Freelancer Overview

Since July 2025 I've worked as a Physics Expert at Turing, doing RLHF for physics at both the undergraduate and PhD level. My work there covers the full pipeline: writing rubrics for text-based and image-based questions, data labelling, and finding "headroom" questions, meaning problems hard enough to challenge state-of-the-art models and expose where their reasoning breaks down. So I'm already comfortable evaluating model outputs rigorously, designing clear grading criteria, and probing for the edges where models fail rather than just checking surface-level correctness. What sets me apart is the depth of physics behind that judgment. I'm a physics PhD student at the University of Toronto with a 3.88 CGPA, graduate coursework in quantum field theory, quantum optics, and condensed matter, and a co-authored paper in quantum optics (arXiv:2409.08230). On top of that I have several years of teaching experience as a TA and tutor in math and physics, plus an olympiad background, which has trained me to explain technical reasoning clearly and to quickly catch where an argument goes subtly wrong. That instinct, spotting answers that look right but aren't, is exactly what matters when evaluating AI-generated physics.

Labeling Experience

Physics Expert

ImageRLHF

I work on RLHF (reinforcement learning from human feedback) to improve frontier large language models on physics, from undergraduate through PhD level. My work spans authoring headroom questions, including frontier-level and image-based items designed to challenge state-of-the-art models and surface where their reasoning breaks down, as well as CUJ (Critical User Journey) tasks. I build rubrics for both text- and image-based physics questions, and I assess LLM outputs across modalities, including text, image, HTML, and video. The core of the role is rigorous, step-level evaluation of physics reasoning: grading not just final-answer correctness but the validity of each step, and identifying solutions that appear correct while concealing a subtle conceptual or mathematical error.

2025 - Present

Education

Work History