Skip to content
#

varolline

Here are 21 public repositories matching this topic...

Hybrid RAG system that combines BM25 keyword and FAISS vector search with cross-encoder reranking and source-cited answers. Includes a 47-question benchmark that measures Hit@k, MRR and answer correctness, and compares fixed-size vs recursive chunking over 22 PDF, DOCX and TXT documents.

  • Updated Oct 27, 2026
  • Python

Sentiment analysis that classifies Amazon product reviews as Positive, Negative or Neutral. TF-IDF with Linear SVM, Logistic Regression and Naive Bayes, tuned and compared, served via FastAPI with a web interface. Accuracy 83.3%, macro F1 0.71.

  • Updated Oct 18, 2026
  • Python

RAG-based PDF Question-Answering chatbot using FAISS, sentence-transformers and Google Gemini, with page-level source citations and hallucination-safe "not found" handling.

  • Updated Oct 20, 2026
  • Python

A RAG chatbot built with LangChain and FAISS that answers questions from your documents with source citations, refuses out-of-scope queries, remembers conversation context, and runs through a FastAPI service and command-line chat.

  • Updated Oct 11, 2026
  • Python

Image classifier for 6 scene classes (buildings, forest, glacier, mountain, sea, street) built with PyTorch transfer learning. Fine-tunes a pretrained ResNet-18 on the Intel Image Classification dataset with data augmentation, reaching 89.6% test accuracy. Includes a predict script and Gradio web demo.

  • Updated Oct 21, 2026
  • Python

Explainable AI and fairness audit on the Adult Income dataset. Trains an XGBoost classifier, explains it globally and locally with SHAP and LIME, measures gender and race gaps with Fairlearn, applies Reweighing, ExponentiatedGradient and ThresholdOptimizer, and reports accuracy vs fairness trade-offs.

  • Updated Nov 8, 2026
  • HTML

End-to-end retail demand forecasting system on Kaggle Store Sales data. Compares a Seasonal Naive baseline, SARIMAX and LightGBM (lag and calendar features) using time-based cross-validation (MAE, RMSE). The best model, LightGBM, is served through a FastAPI REST API.

  • Updated Oct 29, 2026
  • Python

A feedforward neural network built from scratch in pure NumPy: forward/backprop, ReLU/Tanh/Sigmoid/Softmax, mini-batch SGD with momentum and gradient checking. Trained on MNIST (97.98% test accuracy) and verified against an equivalent PyTorch model.

  • Updated Nov 12, 2026
  • Python

Measures how 5 prompting strategies (zero-shot, few-shot, chain-of-thought, retrieval-grounded, strict "say unknown") affect accuracy and hallucination in Gemini on 40 factual questions. Includes code, question set, graded results, comparison table and PDF report.

  • Updated Nov 15, 2026
  • Python

Fine-tuned Qwen2.5-1.5B-Instruct with QLoRA on DialogSum for dialogue summarization. Outperforms the base model on 200 held-out examples (ROUGE-1 +64%, BERTScore F1 0.87 to 0.91). Trained in about 33 min on a free Colab T4. Includes code, training logs, evaluation report and side-by-side examples.

  • Updated Nov 4, 2026
  • Python

End-to-end document AI pipeline that extracts invoice number, date, vendor, line items and total from scanned invoices and receipts using Tesseract OCR and a Gemini vision LLM. Pydantic-validated JSON, confidence-based human-review flags and field-level accuracy evaluation.

  • Updated Nov 8, 2026
  • Python

Transformer self-attention built from scratch in PyTorch: scaled dot-product and multi-head attention without built-in modules, a full transformer block, trained on sequence reversal (99.7% accuracy) with attention heatmaps that show how the model learned to focus.

  • Updated Nov 14, 2026
  • Python

Movie recommendation system on MovieLens 100K: popularity baseline, item-KNN and SVD collaborative filtering, content-based model and a tuned hybrid. Evaluated with Precision@K, Recall@K and NDCG on a per-user time-based split, plus a cold-start strategy for new users. Python, scikit-learn.

  • Updated Nov 8, 2026
  • Python

A controlled PyTorch experiment that deliberately overfits a small MNIST subset, then fixes it one remedy at a time: L1, L2, dropout, early stopping, data augmentation, and simpler architecture. Includes train/validation curves, a comparison table, and an analysis of what worked and why.

  • Updated Nov 14, 2026
  • Python

Multi-step LLM agent (Gemini) that calls tools: web search, calculator, read-only SQL and a sandboxed file reader. Pydantic argument validation, retries with backoff, step limit, loop guard and full JSONL reasoning/tool-call logs. Includes a 13-task test suite and failure analysis.

  • Updated Oct 29, 2026
  • Python

Fraud detection on a highly imbalanced dataset (0.17% fraud). Compares class weights, SMOTE, undersampling and threshold tuning with Logistic Regression, evaluated via precision-recall curves, F1/F2 and PR-AUC. Shows why accuracy misleads, tests label-noise robustness and recommends a high-recall approach. Includes code, plots and report.

  • Updated Nov 15, 2026
  • Python

Find and fix 9 deliberate flaws in a broken scikit-learn ML pipeline: target leakage, duplicate rows, preprocessing leakage, time-ordering bug, unscaled features, class imbalance, test-set tuning and wrong metrics. Includes broken vs fixed code, honest before/after results and a detection report.

  • Updated Nov 14, 2026
  • Python

Production-style ML API for California house price prediction. Built with FastAPI and Docker, with request logging, drift monitoring and Locust load testing (0 failures in 5,048 requests, ~10 ms model inference, 33-40 req/s on one worker).

  • Updated Oct 30, 2026
  • HTML

Add this topic to your repo

To associate your repository with the varolline topic, visit your repo's landing page and select "manage topics."

Learn more