A web application for froth flotation bubble segmentation, allowing users to upload images for segmentation or visualize video segmentation by frame.

Nine projects from the Machine Learning and AI in the Life Sciences lab seminar (FU Berlin) — omics ML, 3D computer vision, and geometric deep learning, each with a written report

Real-time scene text detection and recognition for Arabic storefronts and signage using YOLOv4 and CRNN-CTC

An open-source research project exploring deep learning for skin-lesion image classification, with model evaluation and explainability experiments. Research use only—not a medical diagnostic tool.

A professional, production-grade AI platform for rice crop analysis combining segmentation, disease detection, and variety classification into an intelligent pipeline.

木材横截面树髓定位:传统几何、ArcPith 年轮弧反演、分层评价与子弧解释、2D-OA-Pith 深度学习。基于 UruDendro4 公开数据集(Marichal et al., ICPRS 2025)

AlphaZero-style chess engine in PyTorch: SE-ResNet policy/value net guiding batched MCTS with Syzygy tablebases, pretrained on Lichess games then self-play.

Сравнение полносвязных (MLP) и свёрточных (CNN) нейронных сетей для распознавания рукописных цифр MNIST. Реализация моделей на PyTorch, эксперименты с архитектурами и анализ качества классификации.

Brain-tumor MRI classification, audited: 73% of the test set leaks from train — the famous 99% recipe scores 96.06%

Reversible GANs for image-to-image translation - PyTorch implementation with reversible memory-saving blocks.

Modular, framework-free implementation of a Transformer Language Model from scratch in Python and NumPy. No PyTorch, no TensorFlow.

Classical Longstaff–Schwartz vs. neural-network regression for pricing Bermudan max-call options in 2–20 dimensions, validated against exact and European reference prices (PyTorch)

Handwritten digit recognition on MNIST: MLP vs CNN in PyTorch (99.6% test accuracy)

Character-level language model with a stateful GRU, built with TensorFlow 2 / Keras. Trained on Shakespeare's text to predict the next character and generate new text from a seed string.

Landslide susceptibility, hotspot exposure and rainfall-triggered early warning for Sikkim. FastAPI · rio-tiler · MapLibre

Transformer of "Attention Is All You Need" (Vaswani et al., 2017) from scratch in C++23 dependency-free, header-only: full encoder-decoder, reverse-mode autodiff, training (Adam + noam schedule, label smoothing, dropout) and beam search. Heavily verified (independent reference, finite-difference gradients, 18-bug mutation suite).

34 forecasting algorithms benchmarked on EUR/USD and S&P 500

Protein protein interaction classifier with a frozen ESM2 and a protein disjoint split. Early version 0.1, more coming.

Bengali-language introductory deep learning course covering neural networks, CNNs, and hands-on Python and TensorFlow image-classification lessons.

Production-grade GPU kernels in CUDA C++ and OpenAI Triton for LLMs and Deep Learning, featuring Tiled GEMM, FlashAttention-2, and Fused Ops.

Agentic AI crop disease detection, diagnosis, and agricultural recommendation system using TensorFlow, LangGraph state-based workflows, and FastAPI.

Calibration-free cuffless blood pressure estimation using PPG alone: 1D-CNN + Temporal GRU deep learning, extreme-range isotonic recalibration, split conformal prediction, and real-time reliability triage.

NIFTY 50 market direction prediction using Logistic Regression, FNN, and LSTM with chronological evaluation and backtesting.

Audio AI system for Quran recitation analysis using phoneme recognition, boundary detection, duration estimation, CTC, Wav2Vec2, and Conformer.

MRI görüntülerinden beyin tümörü tespiti ve sınıflandırması: transfer öğrenme (ResNet50, EfficientNetB0, MobileNetV2) + Grad-CAM. Örüntü Tanıma dersi projesi.

AI interview prep coach: webcam emotion detection, voice commands, mock interviews with confidence scoring. FastAPI + React + OpenCV.

Comprendre les Transformers : de « Attention Is All You Need » à un résumé automatique de texte en français avec BARThez, analyse de la cross-attention et évaluation ROUGE / BERTScore.

MLP and CNN classifiers for the SVHN house-number digits dataset, built with TensorFlow 2 (Keras).

Lane detection prototype for ADAS: comparing classical CV (Canny/Hough, IPM + sliding window) vs deep learning (UFLD) on CULane, with full quantitative evaluation and error analysis.

LSTM next-word predictor trained on famous quotes, with a Streamlit UI showing top-K probabilities and sentence continuation.