Designing novel proteins from scratch remains a grand scientific challenge[1]. Here, we developed ElixirSeeker2, an anti-aging peptide design framework based on Safety-by-Design (SbD) principles, assisted by NemaNet, a computer vision-based Caenorhabditis elegans (C. elegans) safety screening platform. ElixirSeeker2 employs the ESM2 language model in a dual-task setup for binary classification of anti-aging peptides and IC50 potency prediction. Results showed that the accuracy of the anti-aging peptide binary classification model reached 84.6%, the AUROC reached 0.93, and the IC50 prediction accuracy reached an R-square of 0.93, which basically met the requirements for external screening. For peptide design, it combines full enumeration for hexapeptides, all possible 64,000,000 (20^6) amino acid combinations and a hybrid generator-based strategy for longer peptides. All designed peptide candidates like ‘SSKQRP’ subsequently underwent functional validation in two aging-related model systems, senescent HEK293 cells and C. elegans organisms. For toxicity and anti-aging effects assessment, we utilized NemaNet, a deep learning model that analyzes over 30,000 annotated synchronized-worm images using natural language processing (NLP)-inspired segmentation to quantify movement patterns, predict biological age[2], and evaluate safety via age deviation metrics. This represents the first de novo pipeline tailored for anti-aging peptides, showcasing an integration of computational biology, artificial intelligence, and classical toxicological models, and offers not only a tool for design but also a window into the hidden language by which proteins encode healing and harm.