Skip to content

Bondu Chandu

Machine Learning Engineer • Hyderabad, Telangana, India • c*****************@gmail.com • +91*******146 • drivetube.ai/•••••

Professional Summary

Machine Learning Engineer with 0 years of experience pursuing M.Tech in Mathematics and Computing. Experienced in building distributed ML pipelines, multivariate time-series forecasting, and communication-efficient distributed training for large transactional and telemetry datasets. Skilled in PyTorch-based model development, feature engineering, model stabilization techniques (AMP, gradient clipping, LR scheduling), and deploying inference endpoints with FastAPI and Docker.

Technical Skills

Programming Languages: Python,C,C++,R
Frameworks and Libraries: NumPy,Pandas,Matplotlib,Seaborn,Scikit-learn,PyTorch,Neural Networks,LSTM,Reinforcement Learning,FastAPI
Databases: SQL,PostgreSQL,MongoDB
Cloud and DevOps: Docker,Linux
Data and Analytics: XGBoost,LightGBM,Feature Engineering,Hyperparameter Tuning
Tools and Methodologies: Git
Distributed Training & Optimization: LocalSGD,Distributed Training,Automatic Mixed Precision,Gradient Clipping,Learning Rate Scheduling,Cosine Annealing
Problem Domains: Time-Series Forecasting,Anomaly,Fraud Detection,Simulation & Stochastic Optimization

Work Experience

ISRO SDSC SHAR
Sriharikota / Hyderabad, India
Machine Learning Intern
Feb 2024 – May 2024
Research internship at an Indian space research center working on atmospheric telemetry and launch-site weather forecasting using time-series ML models.
Tech Stack: Python, PyTorch, Pandas, NumPy, Git, Linux
  • Built an end-to-end LSTM forecasting pipeline for multivariate atmospheric time-series (420K+ observations across 7 variables) including reusable preprocessing, feature engineering, and sequence-generation workflows using Python, Pandas and NumPy.
  • Implemented circular encoding for wind-direction to preserve angular continuity; integrated the encoding into the pipeline and evaluated forecasting impact using R², RMSE and MAE across multiple model configurations with a 70/20/10 train/validation/test split.
  • Performed hyperparameter tuning and training-stabilization techniques (gradient clipping, LR scheduling, AMP) during LSTM training to improve convergence and robustness; achieved R² up to 0.9998 and sub-unit RMSE on most continuous variables.
  • Designed multi-horizon sequence generation and efficient batching to enable GPU-accelerated training on large time-series datasets and to support multi-step forecasting experiments.
  • Established a consistent model evaluation regime to compare configurations across variables and report operational metrics for forecasting decisions used by domain stakeholders.
  • Packaged and documented reproducible experiment code and model artifacts on GitHub to support future deployment and research handoff.

Projects

Distributed Fraud Detection for Payment Systems using LocalSGD
Tools Used: Python, PyTorch, LocalSGD, FastAPI, Docker, AMP
  • Engineered a communication-efficient distributed fraud detection framework using PyTorch LocalSGD across 5 simulated worker nodes with non-IID data partitioning to mimic geographically distributed payment data centers.
  • Trained and evaluated models on a 6.36M-transaction dataset, achieving 99.94% accuracy and 0.9906 ROC-AUC with a 0.680 F1-score on the minority fraud class under realistic class imbalance.
  • Integrated Automatic Mixed Precision, gradient clipping, and learning-rate scheduling to stabilize distributed training and reduce time-to-convergence.
  • Deployed real-time fraud inference via FastAPI inside Docker containers and built monitoring scripts and network simulations to evaluate distributed training behavior under network variability.
RapidoSim — Ride-Hailing Simulation
Tools Used: Python, Reinforcement Learning, Stochastic Optimization
  • Developed a multi-agent ride-hailing simulation modeling supply-demand dynamics as a Markov Decision Process with dynamic pricing as the action space.
  • Applied reinforcement learning and stochastic optimization to benchmark adaptive pricing policies against static and surge-multiplier baselines within a modular experimentation framework.
Thermal Throttling Prediction via System Telemetry
Tools Used: Python, XGBoost, Scikit-learn, Feature Engineering
  • Built an ML pipeline to predict CPU thermal throttling from hardware telemetry, engineering temporal features to capture trends and seasonality.
  • Evaluated XGBoost models for early throttling detection, achieving R² ≈ 0.355 and MAE ≈ 0.73°C to support proactive thermal management decisions.
Convex Optimization Essentials for Linear Models
Tools Used: Python, NumPy, PyTorch, Matplotlib
  • Implemented Gradient Descent, SGD, Mini-Batch GD, and SAGA from scratch and benchmarked optimization algorithms to study convergence properties.
  • Validated optimizer performance with cross-validation experiments, achieving 96.5% cross-validation accuracy and visualized convergence behavior across training settings.

Education

Dr. B.R. Ambedkar National Institute of Technology, Jalandhar
M.Tech, Mathematics and Computing • Jalandhar, India • 2025 – 2027
JNTUA College of Engineering, Kalikiri
B.Tech, Computer Science and Engineering • Kalikiri, India • 2020 – 2024

Powered by Drivetube · Create your own profile at drivetube.ai