Indie hacker

Sanjeeva Rao  • Peerlist

- Built an end-to-end FinOps demo that ingests 90-day AWS Cost-Explorer data (or generates synthetic), transforms to KPI tables in DuckDB/Parquet, and serves insights via a public Streamlit dashboard. - Implemented incremental Polars, DuckDB pipeline that executes in ~1s and cuts storage > 3X versus input (column pruning, Snappy compression). - Designed KPI suite: cost-per-team, idle-compute %, forecast burn, anomaly score (±2 σ). - Automated CI/CD: unit tests and lint run on every commit via GitHub Actions, auto-build, and deploy dashboard.

Cloud-Cost and Usage Observatory

- Used keras from TensorFlow, numpy, and matplotlib in Python to design, implement and evaluate a Deep CNN for speech recognition to achieve an accuracy of>60% on the validation data. - Used 3x3 grid search to systematically investigate the effect of adding layers and the number of filters per layer to the model. - Designed, implemented, and evaluated a model averaging scheme to regularise the network and improve generalization performance.

CNN model for Speech Recognition

Implemented and evaluated scalable trajectory clustering pipelines using Apache Spark for large-scale GPS data from the Microsoft GeoLife dataset, comparing DBSCAN and TRACLUS across spatial and temporal partitions. Developed a PySpark-based pipeline to ingest, parse, segment, and cluster over 10,000 GPS trajectories, applying region-, time-, and user-based partitioning strategies. Applied DBSCAN and a custom TRACLUS implementation using Euclidean and DTW distances, each assessed for clustering quality (Silhouette Score), runtime, and memory usage. Demonstrated that DTW-based TRACLUS produced fewer noise points (6%) but had approximately 130 times longer runtime (569 seconds) than DBSCAN (4.39 seconds), indicating a significant trade-off between accuracy and efficiency. Region-based partitioning enhanced performance by approximately 20%, reducing shuffle overhead and facilitating improved cluster workload distribution in Spark. Utilized Pandas UDFs and task-level tuning for DTW alignment, achieving consistent performance while managing executor memory and avoiding recomputation using Spark's caching mechanisms. Evaluated across six clustering configurations, each repeated three times, and monitored with Spark Web UI to analyze runtime patterns and task skew. Technologies Used : Apache Spark, PySpark, Databricks, Pandas UDF, MLlib,Jupyter Notebooks

Distributed Trajectory Clustering with Apache

Worked in a team of 6, with a budget of 300 GBP to build an autonomous industrial facility using 4 and 6 Degree of Freedom robot manipulators to construct a Lego house. Key Responsibilities – • Was the Head of Integration and Testing alongside the Software and Hardware lead. • Worked on the planning of systems development lifecycle using system engineering principles and represented it in V-model. • Learned ROS as an integration software to program using Python and C++ in Linux(Ubuntu) and used it for design and to assist simulation done using MATLAB. • Led the team in performing and management of 6 Degree of Freedom and 4 Degree of Freedom Robot Manipulators using Raspberry Pi and Wi-Fi for control and communication. • Worked on data gathering from the Robot Manipulators to analyze and improve the efficiency of inverse kinematics for precise end-effector positions. • Used version control technology Git to collaborate with other team members. • Presented technical aspects of the project to an industrial panel constituting of engineering professionals.

Duplo House Builder - Industry 4.0 (Factory o

Developed classification and regression models to predict pathological complete response (PCR) and relapse-free survival (RFS) in breast cancer patients using clinical and MRI-derived features from the I-SPY2 trial dataset. Preprocessed over 100+ features, handling missing values, outliers (2,076 removed), and applying Min-Max normalization and Principal Component Analysis (PCA) to retain 95.64% variance from 13 MRI-based components. Applied SMOTE to balance the PCR classification task (25% positive class), improving sensitivity while mitigating class imbalance. Evaluated 9 ML classifiers (Logistic Regression, SVM, LightGBM, XGBoost, etc.); selected top models based on balanced accuracy, AUC-ROC, and F1 score: Logistic Regression (Balanced Accuracy: 0.6380, AUC-ROC: 0.6565, F1 Score: 0.4314) SVM (Balanced Accuracy: 0.6110, highest sensitivity: 0.7059) Performed deep hyperparameter tuning with elastic net regularization and class weight balancing for optimal generalisation. For RFS (regression task), compared models like Random Forest, SVR, MLP, and Ridge Regression: Final Random Forest Regressor achieved MAE: 19.63, R²: 0.08, showing nonlinear modeling advantage. Tools & Technologies: Python, pandas, scikit-learn, XGBoost, LightGBM, SHAP, NumPy, Matplotlib, SMOTE, PCA.