• Deutsch

Kavyashree Byalya
Nanjegowda

AI Software Engineer | M.Sc. Digital Engineering

• Email
• LinkedIn
• GitHub

I’m an AI/ML focused software engineer with 4+ years of industry experience delivering automation, agentic systems, and production-ready engineering solutions.I’m pursuing an M.Sc. in Digital Engineering at Otto-von-Guericke University Magdeburg, with a focus on Machine Learning, Data Engineering, and intelligent systems.In my current working-student role, I contribute across software development and quality engineering, while designing and integrating AI capabilities, especially AI-driven multi-agent workflows to strengthen product reliability and impact.Previously, I worked nearly 3 years at Accenture as a Software Engineer, where I strengthened my experience in software development and quality engineering by developing features, supporting production releases and escalations, and collaborating cross-functionally in Agile release cycles.I’m currently seeking roles as an AI Engineer, Machine Learning Engineer, or AI Automation Engineer, and I’m also open to Data Scientist positions focused on deploying practical AI in real-world applications.

Projects

These projects showcase my work across AI/ML, Data Analytics, and software engineering where I turn ideas into reliable, working systems. I take projects end-to-end, from problem framing and data preparation to modeling, evaluation, and delivery through interactive dashboards or deployed prototypes. You'll find reproducible ML workflows, explainable results, and clean implementation focused on real-world impact. Each solution is built to solve a specific problem, validated with data, and designed for clarity and usability.

• View projects

Thank you for taking the time to explore my portfolio!
If you think I could be a good fit for your team, I’d love to connect. You can reach me via email or LinkedIn, or use the form below.

Feature Importance in Time Series For Energy Consumption in CNC Machine

Overview
This project was completed in collaboration with the Autonomous Multi-sensor Systems (AMS) group at Otto von Guericke University (OvGU) as part of an AMS lab project, focusing on feature importance in time series for CNC machine energy consumption. The work used real CNC sensor logs and multiple machine learning models together with explainability techniques - Integrated Gradients (IG), WINIT, LIME, and Permutation Importance to identify which operational signals most influence energy usage and support energy-efficient manufacturing.Problem Statement
1. Which process and sensor variables most influence CNC energy-consumption under different machining conditions?
2. How do feature-importance methods differ in accuracy, runtime, and explanation usefulness for industrial time-series data?
3. How can results be made accessible through an interactive interface so engineers can explore explanations across scenarios?Approach and Methodology
1. Dataset Preparation and Preprocessing
- Used Four real-world time-series datasets collected from a DMC2 CNC machine: DMC2ALCP1, DMC2ALCP2, DMC2SCP1, DMC2SCP2 (Aluminum/Steel × CP1/CP2).
- Removed irrelevant/constant columns (e.g., CYCLE, ADBD0, unused POWER channels).
- Dropped 28 columns containing missing values (NaNs).
- Excluded inactive axis features and retained 52 relevant features tied to active axes and spindle.
- Standardized inputs and normalized targets to support stable training.
- Ran two experimental settings: with and without correlation-based filtering (removed highly correlated features) to study redundancy effects. All modeling and explainability runs were executed offline and outputs (metrics, plots, top features) were stored for consistent comparison and later surfaced in the app.2. Feature Importance Techniques Implemented
Integrated Gradients (IG): Gradient-based attributions for differentiable models (applied to FNN/LSTM).
WINIT: Windowed time-series attribution capturing temporal and delayed effects (applied with LSTM and also explored with XGBoost).
LIME: Local surrogate explanations, evaluated across models including LSTM, Random Forest, and XGBoost.
Permutation Importance (PI): Shuffle-based importance measured by performance drop, used with tree models (Decision Tree, Random Forest).3. Evaluation and Comparison
Compared methods using test loss, execution time, and interpretability and stability of top-ranked features across scenarios.4. Streamlit Application
- Packaged the experiment results into an interactive Streamlit dashboard. The UI supports comparisons across dataset scenarios and explainability methods, including plots and ranked feature lists.
- The user can Select a method (IG/WINIT/LIME/PI) and scenario (material/program), then inspect feature rankings and metric trade-offs instantly.Key Findings1. Feature rankings vary across scenarios and techniques, so comparing multiple explanation views is important for industrial datasets.
2. Integrated Gradients with an FNN and correlation filtering provided a strong overall balance between accuracy, runtime, and interpretability across datasets.
3. Signals related to torque/load and control deviations appeared consistently among influential drivers of energy-consumption patterns.How to Use the App1. Open the dashboard link.
2. Choose a technique (IG / WINIT / LIME / Permutation) and a dataset scenario (material/program).
3. Review the ranked features and compare trade-offs using test loss and execution time.Future Scope
The current app displays precomputed feature-importance outputs, but the same setup can be extended for industrial use by adding a CI/CD pipeline and dataset upload. This would automate preprocessing, model evaluation, and explanation generation, turning the dashboard from a static viewer into a scalable, continuously updated analytics tool.

• back

Interactive 3D Rotary Calciner Simulation for Process Engineering Education

Project Overview
This project was developed in collaboration with the Institute of Process Engineering at Otto von Guericke University (OvGU) and focuses on creating an interactive 3D visualization application for process engineering education, replicating the university's pilot plant rotary calciner, enabling students to explore complex industrial equipment and understand the calcination process through immersive digital interaction.Problem Statement
1. How can complex industrial equipment be made accessible for student learning when physical access to pilot plants is limited and traditional lectures lack hands-on experience?
2. What level of detail and interactivity is needed to help students understand both the structural components and operational processes of a rotary calciner?
3. How can the calcination process be visualized dynamically, showing material movement, thermal decomposition stages, and component interactions in real-time?Approach and Methodology
1. Requirements Gathering and Technical Specifications
- Conducted site visits to OvGU's pilot plant calciner, taking detailed measurements and photographs.
- Documented calciner specifications, analyzed thermal decomposition process, and calculated simulation parameters.2.3D Modeling and Asset Creation
- Created precise component models using SolidWorks for complex mechanical parts (motor, chain drive, girth gear with 33 teeth).
- Developed visual models in Blender for drum structure, support rollers, heating coils, and control systems.
- Exported models as .fbx format for Unity integration, ensuring proper scaling and texture mapping.3. Unity Application Development
- Implemented Entity-Component-System (ECS) architecture for modular, scalable code structure.
- Developed core management classes: DialogueManager (22 guided scenes), AnimationManager (particle physics), AudioManager (synchronized narration).
- Created dual interaction modes: Guided learning path with narrative dialogue, Free 3D exploration with mouse/slider controls (zoom, rotation).
- Built ButtonControlScript for consistent navigation (Home, Next, Back, Skip buttons across all scenes).4. Physics-Based Simulation
- Programmed for realistic drum rotation synchronized with motor and girth gear movement.
- Implemented gravity and collision physics for authentic material tumbling inside a rotating drum.5. Interactive Learning Features
- Components Explanation Mode: Individual component highlighting with technical specifications and educational descriptions
Simulation Mode: Full operational cycle showing material loading, drum rotation (RPM-controlled), heating activation, particle movement.6. Software Development Lifecycle
- Applied Agile methodology with iterative development, regular supervisor feedback sessions, and continuous refinement.
- Conducted user testing with engineering students to validate educational effectiveness and usability.Key Takeaways
- Cross-disciplinary collaboration between Computer Science and Chemical Engineering can produce effective tools, combining technical modeling expertise with domain-specific process knowledge.
- Dual interaction paradigms (guided vs. exploratory learning) address different user needs, supporting both structured curriculum delivery and self-directed investigation.
- Design patterns (Singleton, ECS, Decorator) ensure maintainability and extensibility, allowing easy addition of new components, processes, or simulation scenarios in future iterations.
- 3D visualization bridges theory-practice gap in engineering education, enabling students to understand complex equipment operation before encountering physical systems in industry.Future Scope
The current application focuses on batch calcination of aluminum hydroxide and can be extended with variable process parameters allowing users to adjust temperature profiles, rotation speeds, and material properties to observe different calcination outcomes. Multi-material support could include limestone, magnesite, and petroleum coke calcination processes with material-specific thermal behavior. Integration of real-time data visualization showing temperature graphs, moisture content curves, and energy consumption metrics would strengthen quantitative learning. VR/AR deployment would provide truly immersive experiences, while assessment modules with quizzes and simulation challenges could measure learning outcomes and provide automated feedback for self-paced education.

Sales and Profit Dashboard in Power BI

Project Overview
This project, developed as part of the Data Visualization course at the Visualization Group, Otto-von-Guericke University, presents an interactive Power BI dashboard built on a multi-year U.S. superstore sales and profit dataset sourced from Kaggle. It enables users to analyze profitability by region, customer segment, product category, and time, using clear visuals and an intuitive, exploratory layout. The focus is on advanced, interactive reporting that makes it easier to spot performance patterns and business drivers.Problem Statement
1. How can superstore retail data be visualized in a way that clearly surfaces sales and profit insights across customer groups, product categories, and geographic regions?
2. What profit patterns emerge over time, and how do they differ across regions and product sub-categories?
3. Which customer segments drive the most profit, and which categories show sustained revenue performance?
4. How can an interactive dashboard support fast, self-service exploration of complex, multi-dimensional sales data for better decision-making?Methodology and Steps
Dataset:
The selected Kaggle dataset emulates realistic retail transactions, making it suitable for meaningful business analysis and visualization with below features:
- Roughly 9,994 records covering the years 2016–2019.
- All measures are in USD.
- Key fields include customer segment (Consumer, Corporate, Home Office), product categories and sub-categories (Furniture, Office Supplies, Technology), shipping mode, order date, geographic attributes (50 U.S. states grouped into four regions), sales, quantity, and profit.1) Data preparation and cleaning
- Imported the dataset and reviewed structure, data types, and field definitions.
- Performed quality checks on date ranges, categorical consistency, and validity of sales/profit values to ensure reliable reporting.2) Power BI dashboard design
Built a set of visuals to analyze performance from multiple angles, including:
- Pie charts for regional profit share.
- Donut charts for profit split by customer segment and product category.
- Bar charts to compare profit across product sub-categories.
- Line charts to track profit trends by year/quarter.
- Tables for detailed sales and profit figures.
- Slicers (e.g., state) and KPI cards to support quick filtering and high-level performance summaries.
The report is designed to support drill-down and cross-filtering for deeper analysis.3) Interactivity and user experience
- Enabled interactive exploration through slicers, hover tooltips, and visual cross-highlighting.
- Organized the layout to make navigation straightforward and to help users move from overview KPIs to detailed breakdowns.Key Takeaways
1. The dashboard provides a consolidated view of sales and profit across time, geography, and product structure to support data-driven decisions.
2. Multiple visual perspectives help identify strong-performing regions and categories, as well as sub-categories that may be unprofitable.
3. Customer segmentation analysis shows how different buyer groups contribute differently to profitability.
4. Interactivity (filters and drill-down) improves exploratory analysis and enables customized views without rewriting reports.How to view the Dashboard
1. Download the Power BI report (.pbix) from GitHub
2. Open it using Power BI Desktop (free Microsoft application).
3. Use slicers and interactive visuals to explore profit and sales across states, regions, segments, and product lines.