Sai Teja Reddy

Results-oriented Data Scientist with 5+ years of experience delivering actionable insights through advanced statistical analysis and machine learning. Proven success in fraud analytics, forecasting, speech/text analytics, and end-to-end ML pipeline development. Salesforce Certified AI Associate and AI Specialist, with deep expertise in Python, R, SQL, Tableau, and Google Colab. Known for cross-functional collaboration and commitment to ACA’s Guiding Principles.

• Salesforce Certified AI Associate
• Salesforce Certified AI Specialist

• Languages & Tools: Python, R, SQL, Tableau, Power BI, Jupyter, Google Colab, Git
• ML & NLP Libraries: scikit-learn, TensorFlow, PyTorch, Hugging Face Transformers, Librosa
• Techniques: Ensemble Learning, Decision Trees, Random Forests, SVM, Boosted Models, Neural Networks, Time-Series Forecasting, Anomaly Detection, Hyperparameter Tuning, Grid Search CV, RESTful APIs, Model Monitoring

LLM Classification Fine-Tuning — Kaggle Competition

• Fine-tuned a transformer-based Large Language Model (LLM) for multi-class text classification tasks on unstructured datasets.
• Used Google Colab for scalable training, leveraging Hugging Face Transformers and PyTorch to customize pre-trained models (e.g., DistilBERT).
• Applied text preprocessing pipelines: tokenization, padding, attention masks, and managed imbalanced data with stratified sampling.
• Evaluated model performance with F1 Score, Precision, Recall, and built confusion matrices to diagnose classification errors.
• Optimized hyperparameters (learning rate, batch size) for best generalization.
• Documented workflows using Jupyter Notebooks and versioned experiments with Git.

HCU Speech Recognition Challenge 2025 — Kaggle Competition

• Built an end-to-end Automatic Speech Recognition (ASR) system to transcribe noisy speech audio to text.
• Extracted acoustic features like MFCCs, spectrograms, and applied data augmentation to improve model robustness to diverse accents.
• Designed and trained sequence-to-sequence deep learning architectures combining CNNs, RNNs, and attention layers using TensorFlow/Keras in Google Colab.
• Evaluated models using Word Error Rate (WER) and fine-tuned parameters for minimal transcription errors.
• Applied best practices in audio signal processing using Librosa and visualized feature maps for model interpretability

Fraud Analytics Model — Ericsson Canada

• Created an ensemble anomaly detection system combining decision trees and boosting algorithms to identify suspicious transactions in near real-time.
• Automated feature selection and dimensionality reduction to improve model performance and interpretability.
• Integrated model predictions into operational dashboards for fraud monitoring and rapid response

Forecasting Pipeline — Ericsson Canada

• Developed a fully automated time-series forecasting pipeline using Python (ARIMA/Prophet) to predict network traffic and capacity needs.
• Deployed the pipeline on Google Colab and integrated output visualizations in Tableau for stakeholders.
• Implemented model performance tracking using rolling forecasts and KPI dashboards.