Jeffy — Pretrained Classifiers for Text, Games, and More

Pretrained classifiers.
CPU only. pip install.

68 classifiers ship ready to use. Text, game state, numeric features. No GPU, no API keys, no cloud dependency.

Try it live
Classifier:

Latest

🧠❓
← or →

Can AI Read Your Mind?

Pick left or right — a Markov chain learns your pattern and predicts your next move. 7 game modes including Boss Fight, Survival, and a global leaderboard.

7 modes Leaderboard Mobile + keyboard
JeffyBench: 16 models compared on banking77

When LLMs Fail at Classification

16 models benchmarked across 3 tasks. As class count grows, LLMs collapse — Jeffy 94.3% on banking77 vs 68% best LLM.

16 models 3 tasks 94.3% on 77 classes

Live demos

Pac-Man AI playing via text classification

Pac-Man AI

Regex parses junction descriptions into features, Jeffy classifies the best direction. 8,035 mean score, 2.5x the current #1, 3ms per decision.

2.5x #1 3ms / decision
Read the tutorial →
Doom classifier demo: real-time game-state decisions
0%

Doom Classifier

Real-time game-state classification. 24 numeric features extracted each frame decide whether to fire, turn left, or turn right.

3 classes 100% accuracy
Open demo →
Poker AI demo: four players making decisions at a Texas Hold'em table
0%

Poker AI

Four AI players at a Texas Hold'em table. Each makes fold, check, call, or raise decisions from 18 game-state features, streamed live.

4 classes 88% accuracy
Open demo →
Fly Navigation Demo
0%

Fly Navigation

A biomechanical fruit fly forages for food via chemotaxis. Jeffy classifies walk, turn-left, or turn-right from 16 physics features at each step.

3 classes 98% accuracy
Open demo →
68
Classifiers
99.1%
Top accuracy · benchmark →
<80ms
Latency
CPU
No GPU needed
~/project

17 English classifiers

51 multilingual classifiers — 60 voice-command intents per language, frozen multilingual embeddings

Predict

from jeffy.engine import Engine engine = Engine() engine.load() result = engine.predict( "banking77", "I was charged twice for the same transaction" ) # {"label": "transaction_charged_twice", "confidence": 0.999}

Train your own

from jeffy.train import train_classifier clf = train_classifier( texts=["great product!", "terrible", ...], labels=["positive", "negative", ...], task_id="my_reviews", ) clf.predict("exceeded expectations") # {"label": "positive", "confidence": 0.87}