🤖 Edge AI Semantic Mapping — RPi5 + Hailo-8
Real-time 2D semantic mapping on Raspberry Pi 5 with Hailo-8 NPU for assistive navigation. On-device inference at full speed with zero cloud dependency.
Edge AIHailo-8RPi5ROS2ONNX
🐧 Penguin AI — Multilingual RAG + RLHF Chatbot
Production LLM chatbot with 6-layer multilingual RAG pipeline (50+ languages), PPO-style RLHF, agentic ReAct reasoning, and Whisper voice input — powered by Groq + LangChain.
Multilingual RAGRLHFLangChainFAISSGroqWhisper
🎙️ Speech Emotion Recognition
Classifies 7 emotions from speech using MFCC + Mel-spectrograms with a CNN-LSTM architecture. Live mic and file upload support.
Speech AICNN-LSTMMFCCLibROSAPyTorch
🎭 Image Classification & Segmentation
PCA from scratch → ResNet-50 multi-label classifier (20 VOC classes) → semantic segmentation with ResNet-50 + ASPP decoder (mIoU 0.79) → Wilcoxon signed-rank + Bootstrap CI statistical testing.
PyTorchResNet-50ASPPU-NetPCAWilcoxon
🔢 Effect of Precision on PDE Solvers
FP64 vs FP32 vs FP16 across 4 finite-difference schemes (Godunov, Lax-Friedrichs, Lax-Wendroff, MUSCL) on advection-diffusion + Burgers' equations. FP16 loses stability on high-Re grids; MUSCL + SSP-RK2 preserves TVD at FP32.
PyTorchPDEsFinite DifferenceNumerical MethodsFP16/32/64
📉 Gradient Optimization — From Scratch
NumPy implementations of 7 optimizers (GD, SGD, Momentum, Nesterov, AdaGrad, RMSProp, Adam). Empirically verified O(1/k²) Nesterov convergence, condition number theory, and Adam generalization gap on convex + non-convex landscapes.
NumPyOptimizationConvex AnalysisAdamNesterov
🎸 Multi-Modal Instrument Recommender
Recommends instruments from audio, image, and text inputs using multimodal embeddings (CLIP + audio features) and cosine similarity search.
MultimodalCLIPEmbeddingsAudio AICosine Search
🏃 Human Activity Recognition
Classifies 6 activities from smartphone accelerometer + gyroscope using 5 ML models (SVM 95.5%). Real-time continuous detection on mobile — LinearSVC runs fully in-browser, no server needed.
SVMRandom ForestXGBoostLightGBMScikit-learnStreamlit
💳 Credit Card Fraud Detection
Anomaly detection on highly imbalanced transaction data using XGBoost + SMOTE. Achieves 99%+ precision with real-time scoring.
XGBoostSMOTEAnomaly DetectionScikit-learn
😷 Face Mask Detector — 3-Class
Compared EfficientNetB0 TFLite FP16, MobileNetV2, and YOLOv8 for real-time 3-class mask detection (Mask On / Incorrect / No Mask). EfficientNetB0 achieved 98.33% accuracy. Pipelined with OpenCV SSD face detector and deployed on Streamlit Cloud.
EfficientNetB0YOLOv8MobileNetV2OpenCVTFLiteStreamlit
📈 Stock Market Prediction (LSTM)
LSTM-based forecasting with RSI, MACD, Bollinger Bands as engineered features. Live data via yfinance with interactive Plotly charts.
LSTMTime SeriesyfinancePlotlyPyTorch
🛒 Big Mart Sales Prediction
Stacked ensemble (XGBoost + LightGBM + Ridge) with log-target + target encoding. Lifted R² to 0.63 across 8,500 stores.
XGBoostLightGBMStackingFeature Engineering
🎬 Movie Recommender System
Content-based recommendation using TF-IDF + cosine similarity on 5,000 movies. Fast, interpretable, and accurate suggestions.
TF-IDFCosine SimilarityNLPPandas
