Vishnu N J

Data Science Intern specializing in analytical methodologies and productivity. Deep expertise in machine learning, statistical analysis, and data visualization enhances problem-solving capabilities. Effective communicator adept at converting data insights into strategic business actions.

Emotion classification in EEG data using frequency-aware transformers

• Tools:python,pandas,numpy,tensorflow,sklearn,scipy,matplotlib,seaborn,tranformers
• Engineered an end to end pipeline to classify nine emotional states from 32nchannel EEG signals (FACED
  dataset, 123 subjects).
• Employed Zero Time Windowing and Discrete Fourier Transform to derive theta, alpha, beta & gamma band
  energies.
• Designed a Frequency Aware Transformer (FAT) model capturing cross channel spectral dependencies;
  implemented in PyTorch.
• Achieved 91.2 % average accuracy (F1 ≈ 0.90) in cross subject evaluation, outperforming RNN and vanilla
  transformer baselines by >10 %

Parameter Efficient Fine Tuning (LoRA) for Movie Genre Classification:

• Tools: Python, transformer, peft, sklearn, pytorch, numpy, pandas, kaggleGPU
• Reimplemented low-rank adaptation (LoRA) to fine-tune a 110M parameter BERT base encoder on a multi-label movie synopsis dataset (24k films, 20 genres) with only ~2M trainable parameters
• Achieved micro F1 = 0.81 and macro F1 = 0.77, matching full fine-tuning while reducing GPU memory by 60% and training time by 35%
• Delivered a reproducible notebook showing LoRA rank sweep, genre-wise confusion matrix, and parameter savings visualization

Time Series Anomaly Detection using LSTM Autoencoder:

• Tools: Python, TensorFlow, NumPy, pandas, scikit-learn, Matplotlib, Seaborn, Jupyter Notebook
• Developed an LSTM-based autoencoder to detect anomalies in stock price time series data, focusing on IndusInd Bank stock. The model learns the normal behavior of stock price sequences (Open, High, Low, Close, VWAP, Volume) and flags significant deviations as anomalies based on reconstruction error thresholds.
• Trained the LSTM autoencoder on non-anomalous stock data, calculated reconstruction errors on test sequences, and applied a dynamic threshold (95th percentile of MAE) to identify anomalous points. Detected 46 high-error points, indicating potential market irregularities.
• Delivered a fully reproducible notebook demonstrating preprocessing, sequence generation, model training, anomaly thresholding, and visualization of reconstruction errors with anomaly markers.

• Club head Anantham, mathematics club of the college
• Head Student Coordinator, Team Mathematics, minor category winner, overall category runner-up in the 12th National Level Techfest Anokha 2024
• Technical head, Anokha for technical events and workshop