Production engineering, pointed at machine learning.

PSPI-T: a transformer approach to prokaryotic small protein identification

Small proteins (≤100 amino acids) do real biological work but are systematically missed by conventional gene finders. The published state of the art, PSPI, used an LSTM plus hand-engineered (n, k)-mer frequency features. I replaced both with embeddings from ESM-2, a pretrained protein language model, and a lightweight classifier — removing the manual feature engineering entirely.

Frozen embeddings with nothing but logistic regression on top matched the published PSPI baseline (AUROC 0.995, F1 0.973). Fine-tuning the encoder beat it: AUROC 0.9978, F1 0.9807.

Robustness of LLMs under paraphrase, noise, and distraction

Most robustness studies perturb inputs one way at a time. I tested three classes at once — paraphrasing, typo corruption, and random distractor insertion — against two 7B instruction-tuned models (DeepSeek-7B-Chat, Mistral-7B-Instruct) on question answering and summarization, 500 examples per task and three seeds per perturbation, measuring both performance drop and across-seed variance.

Paraphrasing dominated, costing up to 26.6% on QA while typos and distractors barely registered. The useful result: clean accuracy did not predict robustness — Mistral scored higher than DeepSeek on clean QA and degraded more. The perturbation that hurt most was also the least stable across seeds.

Bearing fault diagnosis on CWRU vibration data

Predictive maintenance for rotating machinery: classify bearing condition from 12 kHz drive-end vibration signals in the CWRU benchmark. Statistical and FFT-based features feed classical models (logistic regression, random forest, SVM), benchmarked against deep models on the raw sequence, with confusion matrices, ROC curves, and feature-importance analysis.