Recreated PaliGemma from scratch to deeply understand vision-language model components like SigLIP, Gemma, KV cache, attention types, contrastive learning, rotary embeddings, and multimodal inference architecture.
Key Skills: Vision-Language Models, Contrastive Learning, Rotary Embeddings, KV Cache, Multimodal Inference, PyTorch, Transformer Architectures, SigLIP, Gemma
Built an AI-powered recruitment platform with real-time chat, dynamic workspaces, and agentic backend logic using Flask and Gemini API, enhancing recruiter productivity and natural language interaction.
Key Skills: Python, Flask, SQLAlchemy, Google Gemini API, React, TypeScript, Prompt Engineering, Flask-SocketIO, PostgreSQL, Socket.IO
Developed an emotion-aware music retrieval system that interprets user context using Gemini Flash, transforming multi-modal inputs into MPNet embeddings for semantic song matching via Pinecone and Spotify.
Key Skills: Multi-modal Search, Generative AI, LLM, Vector Search, MPNet, CLAP, Pinecone, Supabase, API Integration, Streamlit
Designed Omni, a context-aware multimodal AI assistant that recalls memories, assists tasks, navigates spaces, supports learning, and enhances wellbeing through adaptive modes and intelligent sensory integration.
Key Skills: Multimodal AI, Vector Databases, Speech Recognition, RAG, Depth Estimation, Context Awareness, Task Automation, Gemini Integration
Optimized Vision Transformer training for pedestrian intention recognition using LoRA adapters, achieving 90% accuracy with only 0.68% trainable parameters. Leveraged YOLOv8 to extract ROIs and Vision Transformer for classification on JAAD video clips.
Key Skills: Vision Transformer, Parameter Efficient Fine-Tuning, LoRA, YOLOv8, Object Detection, Low-Rank Adaptation, Deep Learning, Data Augmentation
Simulated dynamic medicine delivery using Deep RL, implementing and comparing six algorithms for path optimization and obstacle avoidance, including PPO, A2C, and Deep Q-Networks in custom Gym environments.
Key Skills: Deep Reinforcement Learning, Q-Learning, Deep Q Networks, A2C, PPO, Actor-Critic, Path Planning, Obstacle Avoidance, Gymnasium, Python
Engineered a smart meal planning pipeline using object detection on food images and recipe tokenization, enabling personalized, nutrition-rich suggestions powered by BART and Hugging Face models.
Key Skills: Multimodal AI, Text Generation, Tokenization, Deep Learning, Hugging Face, Grounding DINO, BART, T5, BERT-FDA, Recipe Generation, Named Entity Recognition
Integrated multiple data modalities to predict crime types and trends using Random Forest, XGBoost, LSTM, and LDA topic modeling, enabling strategic identification of high-risk areas.
Key Skills: Machine Learning, LSTM, Random Forest, XGBoost, Bagging, Boosting, Topic Modeling, LDA, Clustering, BIRCH, Mini Batch KMeans, Time-Series Analysis
Applied attention mechanisms and UGTN transformer models for high-dimensional feature extraction, enhancing pneumonia classification accuracy using COVID-19 Radiography images with TensorFlow and PyTorch frameworks.
Key Skills: Attention Mechanisms, Transformers, UGTN, TensorFlow, PyTorch, Keras, OpenCV, Python, Medical Imaging, Deep Learning
Created a high-performance PostgreSQL database system for Formula 1 analytics, focusing on query cost optimization and indexing to support complex race strategy queries.
Key Skills: SQL, PostgreSQL, Query Optimization, Indexing, Data Management, Python, HTML, CSS, Database Design, Performance Tuning
Designed a predictive policing tool using 14 ML models, LDA topic modeling, and clustering to analyze crime patterns and assist law enforcement with strategic resource planning.
Key Skills: Machine Learning, Neural Networks, LDA, Clustering, Flask, JavaScript, HTML, CSS, Data Visualization, Predictive Modeling, Crime Analysis
Created an image analysis method combining GLCM and self-updating fuzzy clustering for detailed text feature extraction and accurate key-element identification.
Key Skills: Gray-Level Co-occurrence Matrix (GLCM), Fuzzy Clustering, Textual Feature Extraction, Image Analysis, Machine Learning, Patent
Deployed UAV-based monitoring using YOLOv5 and night-vision cameras to detect dead fish with 88% accuracy, enabling real-time alerts and enhancing aquaculture farm health under challenging conditions.
Key Skills: UAV Surveillance, Deep Learning, IoT, Real-Time Monitoring, Night Vision, Alert Systems, Computer Vision, Drone Technology
Created an AI-driven traffic control solution that monitors traffic density and environmental factors in real-time to improve flow and prioritize emergency response.
Key Skills: Deep Learning, Real-Time Video Analysis, Traffic Management, Signal Optimization, Emergency Vehicle Detection, Weather Adaptation, AI, Workflow Efficiency, Patent
Conducted concept-based sentiment analysis on immigration-related tweets using SenticNet APIs and Word2Vec embeddings, benchmarking TextBlob, VADER, and Gensim models under NTU research guidance.
Key Skills: Word2Vec, Semantic Similarity, TextBlob, VADER, Gensim, Scikit-learn, Unsupervised Learning, Sentiment Analysis, Concept Parsing
Surveyed the evolution of image captioning from early methods to advanced deep learning, highlighting LSTM, CNN, attention mechanisms, and transfer learning applied on the Flickr8K dataset.
Key Skills: Image Captioning, Neural Networks, LSTM, CNN, Attention Mechanism, Transfer Learning, NLP, Deep Learning, Flickr8K Dataset
Created a low-cost, real-time snake detection and alert system integrating YOLOv5 with IoT devices, tailored for rural Indian environments and challenging lighting.
Key Skills: Object Detection, IoT Integration, Data Augmentation, Real-Time Detection, Embedded Systems, Alert Mechanisms, Low-Light Vision, Patent
Designed and patented a non-touch digital writing pad using motion detection and OpenCV, offering user-friendly writing tools and export options for education in low-resource settings.
Key Skills: Python, OpenCV, Motion Sensing, Digital Writing Pad, Autocorrect, Spell Check, File Export, Accessibility Technology, Patent
