Prompt Engineering & SFT Response Curation for EdTech AI Assistant
Text
Prompt + Response Writing (SFT)I worked on a supervised fine-tuning dataset for a conversational AI tutor designed to help middle and high school students with math and science homework. My role involved writing high-quality, age-appropriate responses to hundreds of student prompts—everything from "explain quadratic equations like I'm 12" to walking through step-by-step physics problem-solving. The project scope covered roughly 2,400 prompt-response pairs over five months. I handled the full pipeline: drafting initial responses, refining them for tone (encouraging but not overly casual), and running self-consistency checks to make sure explanations didn’t contradict earlier reasoning. I also flagged edge cases where the model might hallucinate formulas or skip logical steps. Quality-wise, I followed the client’s rubric pretty strictly—every response had to be factually accurate, pedagogically sound, and under 150 words unless the prompt specifically asked for depth. I maintained a 98% first-pass acceptance rate, and the few rejections I got were mostly for minor tone adjustments rather than factual errors.