I built a multi-task NLP system to do summarization, emotion detection, and topic classification from one shared encoder as an undergraduate research project, implementing the transformer from scratch and training it on my personal 4070. In April 2026 I was preparing it for submission when I audited my own draft and found that several baseline and ablation numbers were not good enough. I withdrew the paper and did a writeup on Substack. The verified results are there. What I took from it is that the evaluation was the hard part, not the model. Getting honest numbers out of a 28-class multi-label problem with a long tail where threshold choice, split reuse, and metric selection can each move the headline by ten points was harder than getting the training work for the model, and it's where the original draft went wrong. I'm now rebuilding it around a question I think is more useful: given a fixed post-training compute budget, is it better to jointly train one multi-task model, merge separately-tuned experts, or distill from a frontier teacher?

A product I'm building: a structured, non-skippable path through 157+ ML and SWE interview topics, grounded in the candidate's own resume. Two tracks (ML/Research, 71 topics; SWE, 86 topics), a six-stage flow that progresses from big picture → intuition → math/internals → implementation → interview questions → retrieval check. Every session injects the user's real projects and STAR stories into explanations and comprehension checks, so prep is calibrated to engineers who've already shipped code, not beginners.