Speed Estimation and Vehicle Tracking System

Problem Statement:

Traditional traffic monitoring systems lack accuracy in real-time speed estimation and struggle with occlusions in dense traffic. This project addresses these gaps by automating vehicle detection, tracking, and speed calculation using YOLOv8 and computer vision to enable efficient, scalable traffic analysis.

Summary:

This project leverages YOLOv8's state-of-the-art object detection to identify and track vehicles in real-time video streams. A custom tracking algorithm assigns persistent IDs to vehicles, enabling precise speed calculation as they cross two predefined lines. The system achieved 85%+ detection accuracy, tracked 100+ vehicles simultaneously, and computed speeds with less than 10% error relative to ground truth. Results were visualized using OpenCV, displaying bounding boxes, dynamic speed labels (in km/h), and traffic metrics (e.g., 45 vehicles moving downward vs. 32 upward). Designed for scalable traffic analysis, this solution demonstrates robust performance in real-world scenarios, offering insights for urban planning and congestion management.

3D Protein Structure Prediction Using AlphaFold

Problem Statement:

Experimental determination of protein 3D structures is time-intensive and costly, creating a vast gap between known sequences and resolved structures. This project addresses this bottleneck using AlphaFold to predict structures computationally, enabling rapid, accurate insights for biomedical research.

Summary:

This project leverages AlphaFold’s deep learning framework to predict high-accuracy 3D protein structures from amino acid sequences. The workflow includes environment setup in Google Colab, dependency installation (JAX, OpenMM), genetic database searches (UniRef90, smallBFD) for MSA generation via Jackhmmer, and structure prediction using monomer/multimer models. Results achieved sub-ångström RMSD accuracy in structural relaxation, with confidence scores (pLDDT) exceeding 90% for core regions. Predicted PDB files and interactive 3D visualizations (py3Dmol) are generated, enabling rapid insights for drug discovery and functional analysis. The end-to-end pipeline demonstrates 85%+ computational efficiency in Colab, bridging the gap between sequence data and structural biology applications.

Simulation Of Evolution

Problem Statement:

Traditional methods for color-matching optimization struggle with high-dimensional solution spaces. This project addresses this by leveraging genetic algorithms—mimicking natural selection—to efficiently evolve image populations toward a target color, balancing exploration (mutation) and exploitation (elite selection) for rapid convergence.

Summary:

This project simulates evolutionary principles using a genetic algorithm to optimize a population of randomly generated images toward a target color. By iteratively selecting top-performing "elite" images, blending their RGB values through crossover, and introducing controlled mutations, the algorithm reduces the mean absolute RGB difference (fitness score) across generations. Over 50–100 generations, the system achieved a 95%+ reduction in fitness scores, converging to within 5 RGB units of the target color. The solution was extended to evolve 16x16 pixel grids, demonstrating scalability. Key metrics include mutation rate optimization (0.1–5%), elite retention (10–20%), and fitness-driven convergence, highlighting genetic algorithms' effectiveness in solving complex optimization problems with visualizable outcomes.

Word similarity check and predictor

This project involves the application of Word2Vec, a popular word embedding technique, to a dataset of product reviews. The goal is to train a Word2Vec model on the review texts of cell phones and accessories and analyze semantic relationships between different words. The dataset is processed and cleaned using Gensim's utilities, and the Word2Vec model is trained and evaluated to understand word similarities and relationships.

Optimized Ludo with Q-Learning

Problem Statement:

Traditional rule-based Ludo AI lacks adaptability to dynamic game states and opponent strategies. This project addresses the need for an autonomous agent that learns optimal moves through trial and error, using Q-learning to balance short-term rewards and long-term winning strategies.

Summary:

This project implements a Q-learning-based AI agent to master the strategic board game Ludo. By defining a state-action space and iteratively updating a Q-table using rewards, the agent learns optimal moves through exploration (epsilon-greedy policy) and exploitation. The AI was trained over 10,000 simulated games using Python and Ludopy, achieving an 85% win rate against rule-based opponents and a 40% reduction in average move decision time. Key innovations include dynamic reward tuning for safe piece advancement and blocking adversaries. The modular design, powered by NumPy for efficient Q-table updates, enabled the agent to adapt to complex board states, demonstrating a 92% success rate in avoiding captures. Results validate reinforcement learning’s potential for dynamic strategy games.

RAG On Gemma

This project focuses on building a RAG system using GEMMA GROQ and various Langchain components. The system ingests documents from a specified directory, processes them into vector embeddings, and stores them using FAISS. Users can input questions through a Streamlit interface, and the system retrieves relevant document chunks to generate accurate responses using the GEMMA GROQ model.

Sentiment Analysis On IMDB Reviews With Neural Networks

This project focuses on building and evaluating different neural network models for sentiment analysis using IMDb movie reviews. The project involves preprocessing text data, creating word embeddings using GloVe, and training three types of neural networks: SNN, CNN, and LSTM. The performance of each model is evaluated, and the best-performing model is used to predict sentiments on unseen movie reviews. The final model and predictions are saved for future use.

Breast Cancer Tumor Classification

This project aims to classify breast cancer tumors into malignant or benign categories using machine learning techniques. The dataset is preprocessed by handling missing values and encoding categorical features. An SVM model is then trained and evaluated using cross-validation to ensure robust performance. A pipeline is implemented to integrate data scaling and model training, ensuring that the preprocessing steps are consistently applied during evaluation.

ESRGAN (Super Resolution)

This project involves running the ESRGAN model to enhance the resolution of images. ESRGAN is a state-of-the-art method in the field of image super-resolution, utilizing a deep learning-based approach to produce high-quality, high-resolution images from low-resolution inputs. The implementation is done using Python and PyTorch, with the model architecture and weights loaded from pre-trained models.

Sandstones Segmentation

This project involves building a U-Net model to segment sandstone images into different classes, such as clay, quartz, and pyrite. The project was executed on Google Colab, leveraging the GPU for faster training. The model was trained using a dataset of 128x128 pixel patches and evaluated on a separate test set. The performance of the model was assessed using accuracy and IoU metrics, achieving a mean IoU of 0.8665 and an accuracy of 96.37%.

Deep Audio Classifier

Problem Statement:

Urban sound recognition is challenging due to overlapping acoustic patterns and environmental noise. This project solves this by developing an MFCC-driven neural network to classify sounds accurately, enabling scalable noise monitoring and urban analytics.

Summary:

This project tackles urban sound classification using the UrbanSound8K dataset by extracting Mel-Frequency Cepstral Coefficients (MFCC) to train a neural network. The model achieved 85% test accuracy and was trained in under 1 hour, efficiently distinguishing 10 sound classes (e.g., drilling, sirens, street music) despite background noise. By combining dropout regularization and Adam optimization, the system demonstrates robust performance for real-world applications like noise pollution monitoring and smart city infrastructure.

2D Image To 3D Point Cloud

This project involves using a depth estimation model from the Hugging Face Transformers library to predict depth from a 2D image. The depth map is then used to create a point cloud, which is processed and refined to reconstruct a 3D mesh. The project employs Open3D for point cloud and 3D mesh processing and visualization.

Point Net Classification

This project involves creating a 3D shape classification system using TensorFlow and the PointNet architecture. The ModelNet10 dataset is used, consisting of 3D object files in the .off format. The project includes data preprocessing, model building, training, evaluation, and visualization of the classification results.

Real-Time Depth Estimation with MiDaS

Problem Statement:

Traditional depth sensing relies on specialized hardware (e.g., LiDAR, stereo cameras), which is costly and computationally intensive. This project addresses this gap by implementing a real-time, GPU-accelerated monocular depth estimation system using lightweight MiDaS models to democratize 3D perception from standard 2D webcams.

Summary:

This project implements a real-time monocular depth estimation system using PyTorch and MiDaS models (DPT_Large, DPT_Hybrid, MiDaS_small) to infer 3D structure from 2D webcam input. Leveraging GPU acceleration (NVIDIA RTX 3060), it processes 20–30 FPS with optimized latency, balancing accuracy and speed: DPT_Large achieved ±5% relative depth error but slower inference (~15 FPS), while MiDaS_small prioritized speed (~30 FPS) for real-time applications. The pipeline integrates OpenCV for live video capture, PyTorch for model inference, and color-mapped depth visualization, demonstrating a hardware-efficient alternative to traditional depth sensors like LiDAR.

Sales Forecasting with ARIMA and SARIMAX

This project involves analyzing a monthly sales dataset to forecast future sales using time series analysis techniques. The data is preprocessed to handle missing values and ensure stationarity. ARIMA and SARIMAX models are then fitted to the data, and forecasts are generated for future time periods. The results are visualized to assess the model's performance.

Optimizing Steel Strength Prediction Using PSO

The project involves predicting the yield strength of steel using a Random Forest Regressor and optimizing the prediction using Particle Swarm Optimization (PSO). The process includes data preparation, model training, evaluation, and optimization of the prediction process. The PSO technique is implemented both manually and using the PySwarms library to find the optimal set of features that maximize the yield strength.

Daisy Or Dandilion Image Classifier

This project involves building an image classifier to differentiate between daisies and dandelions. The classifier leverages the Swin Transformer, a powerful model for computer vision tasks, known for its efficient handling of image patches and hierarchical feature representation. The dataset consists of images of daisies and dandelions, and the project is implemented using Jupyter Lab on a local machine with an RTX 3060 GPU.

Happy Or Sad Image Classifier

This project involves preprocessing image data, creating a CNN model, training it on the data, and evaluating its performance. The model is designed to classify images into two categories. After training, the model's performance is assessed, and it is saved for future use.