Show HN: Local pretrained classifiers, GPU not needed

Jeffy is a lightweight toolkit of pretrained text classifiers that runs and retrains on CPU without a GPU. It packages 13 task-specific logistic-regression heads that share a single embedding encoder (bge-large-en-v1.5, ~1.2 GB); heads are stored as derived model parameters (numpy .npz) and loaded into a small runtime (~2 GB). Installation and use are simple: install the package, run jeffy-serve to get an HTTP endpoint or use the Python Engine/SDK, and the encoder is downloaded on first use. Training and serving custom models from CSV/TSV/JSONL is supported via CLI or Python, with tuning options (regularization C, test_size, cv_folds) and fast CPU inference (embedding ~50-80 ms, classifier <1 ms).

Performance and provenance are explicit: example tasks include sms_spam (99.1% test accuracy), imdb (94.8%), banking77 (77 classes, 94.3%), ag_news (90.5%), while some tasks underperform compared to task-specific models (SNLI 65.6%, tweet_eval_sentiment 66.2%; emotion has limited class coverage). Probabilities are uncalibrated. Reproducible builds can retrain all heads (~40 minutes, ~5 GB download). Security measures include integrity hashes in per-capability manifests and a warning to only load custom pickles from trusted sources. Code is MIT-licensed; dataset licenses vary and redistribution permissions are not independently verified.