A 272M-parameter encoder-decoder transformer I implemented from scratch (Pre-LN, RMSNorm, T5 relative position bias, FlashAttention, gated-GELU FFN, KV-cache) with explicit FLAN-T5-base weight mapping — jointly trained on abstractive summarization, topic classification, and multi-label emotion detection. I diagnosed negative transfer in naive MTL and designed two targeted interventions: learned attention pooling for classification heads and temperature-based task sampling (α=0.5). Emotion-detection sample-averaged F1 went 0.199 → 0.352 (+77%), beating the single-task baseline; per-class threshold tuning lifted macro F1 to 0.290. Shared-encoder MTL also beats a BERT-base topic-classification specialist (85.7% vs. 83.1%). 5-seed validation, bootstrap CIs, cross-task MD5 dedup, gradient-conflict diagnostics, MLflow tracking. Targeting arXiv + EMNLP 2026.

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.