A professional AI learning platform showcasing Machine Learning, Deep Learning, NLP, and Agentic AI.
From theory to production — built and self-hosted on a Synology NAS.
One product across the whole stack: a daily data pipeline over 22,000+ men's and women's matches, a statistics API and an ODI World Cup 2027 predictor are live now; machine-learning and deep-learning challengers and an LLM analyst with published accuracy scores come next.
Each project is a containerized microservice — built, tested and served from a Synology NAS.
cricstat Elo ratings and 50,000 simulated tournaments give every team's World Cup 2027 chances, backtested on the 2019 and 2023 World Cups. ML and DL challengers next.
/cricket/predictor/ Open the predictor → How it works in cricstatcrıcstat →K-Nearest Neighbors classifier on the classic Iris dataset. Submit features, get classification with confidence scores.
/ml/iris-knn/ Try it →Linear Regression on the California Housing dataset. Predict median house values from 8 features with an interactive slider UI.
/ml/housing-linear/ Try it →PyCaret AutoML pipeline that benchmarks 7 algorithms on the Titanic survival dataset and serves the best model with AUC/F1/Accuracy metrics.
/ml/titanic-automl/ Try it →Live win probability and player-strength ratings for the predictor, learned from 11.6M deliveries.
/cricket/ How it works in cricstatcrıcstat →PyTorch CNN classifier on MNIST. Draw a digit on the HTML5 canvas, get the prediction with per-class confidence bars. ≥98% test accuracy.
/dl/mnist-cnn/ Try it →PyTorch LSTM trained on 2.5 years of NYC CitiBike daily ride counts. Pick a date, get a 14-day forecast with an MC-Dropout confidence band and a compare-to-actuals overlay.
/dl/lstm-forecast/ Try it →Real-time object detection — bounding boxes and class labels via API.
/dl/object-detection PlannedData-to-text match summaries with every claim verified against the statistics.
/cricket/ How it works in cricstatcrıcstat →Are two Quora questions asking the same thing? Bag-of-words + 22 engineered features (token, length, fuzzy) → Random Forest, with a probability slider and a “why” panel.
/nlp/quora-randomforest/ Try it →Paste any text and watch every step — cleaning, tokenization, stemming vs. lemmatization — then compare one-hot, Bag of Words, n-grams and TF-IDF vectors side by side, each scored by the same classifier on 25,000 IMDB reviews.
/nlp/imdb-textrep/ Try it →CBOW vs skip-gram vs fastText trained on 17M words of Wikipedia, next to GloVe 6B: nearest words, king − man + woman analogies, similarity and a 2-D PCA map.
/nlp/text8-word2vec/ Try it →Tags each word with a hidden Markov model counted by hand from 200k words of web English, drawing the Viterbi trellis and best path — side by side with spaCy's neural tagger (94.4% vs 95.1%).
/nlp/ewt-hmm/ Try it →LangGraph agent with typed tools over the stats API and published evaluation scores.
/cricket/ How it works in cricstatcrıcstat →Autonomous AI agent powered by Claude API. Multi-step reasoning, tool use, and structured outputs.
/agentic/claude-agent PlannedProduction-grade tools across all domains.
Shared services powering all ML experiments and model tracking.
Experiment tracking, run history, parameters, metrics, and model artifact storage for all ML projects.
Open MLflow →S3-compatible object storage for model artifacts, datasets, and MLflow experiment outputs.
ml-minio:9000 · internalShared relational database for prediction audit trails and structured ML project data.
ml-postgres:5432 · internalIn-memory cache for inference results, rate limiting, and session state across ML services.
ml-redis:6379 · internal