Hitesh Shanmukha

Code-Switched ASR + Zero-Shot Voice Cloning

Code-switched lecture transcription and cross-lingual voice cloning pipeline. Fine-tunes speech foundation models for Indian-English classroom audio and generates zero-shot speaker-preserving speech in multiple languages.

• PyTorch
• Whisper
• XTTS
• Speech Processing

Neural Machine Translation (Seq2Seq + Attention)

Encoder-decoder LSTM with Bahdanau attention for German→English and Hindi→English. Trained on 29K DE–EN and 10K HI–EN pairs. Applied scheduled teacher forcing, gradient clipping, and checkpointing over 100 epochs.

BLEU 36.05 on DE→English · validated on 500+ test sentences

• PyTorch
• TorchText
• SpaCy
• LSTM
• Attention

Multi-Object Tracking on MOT17

Three tracking-by-detection pipelines benchmarked on MOT17. Built Faster R-CNN + SORT with anchor tuning, focal loss, and adaptive Kalman filtering; integrated YOLOv8 + DeepSORT for robust ID preservation.

MOTA 0.529 · IDF1 0.578

• PyTorch
• OpenCV
• YOLOv5/v8
• Faster R-CNN
• DeepSORT
• Hungarian

Marketing Research AI Agent

Autonomous agent that runs end-to-end market research — topic discovery, web search, synthesis, and report generation — orchestrated with LLM tool-use and structured output.

• Python
• LangChain
• LLM Tool-Use
• Agents

Brain Tumour Detection

CNN-based classifier for brain MRI scans. Compared transfer-learning backbones and data-augmentation strategies on a multi-class tumour dataset.

• TensorFlow/Keras
• CNN
• Transfer Learning
• Medical Imaging

Real-Time Traffic Monitoring

Real-time traffic monitoring system using object detection and tracking on live video. Flags congestion, counts vehicles, and surfaces lane-level analytics.

• Python
• OpenCV
• YOLO
• Real-Time

COVID-19 Chest X-ray Analysis

Preprocessed 9,535 chest X-rays with PCA and lung segmentation. Benchmarked classical ML (Random Forest, XGBoost, SVM) against deep CNNs (VGG-19, ResNet-50, EfficientNet-B3).

ML 93% · DL 97.5% accuracy

• TensorFlow
• Keras
• Scikit-learn
• OpenCV
• ResNet-50

Low-Light Image Enhancement

Deep-learning pipeline for enhancing dark, noisy photographs. Explored retinex-based and CNN-based approaches and compared PSNR / SSIM across methods.

• PyTorch
• CNN
• Image Processing
• PSNR/SSIM

Stock Price Prediction — LSTM

Sequence model for short-term stock price forecasting. Engineered technical indicators as features and evaluated LSTM vs. baseline regressors on multiple tickers.

• PyTorch
• LSTM
• Pandas
• Time-Series