M.Sc. Computer Science at Blekinge Institute of Technology, Sweden
Curious by nature. Driven to understand problems and build useful solutions.
I'm Pavan Kumar Goud Amikula. I enjoy taking a complex question, breaking it into testable steps, and turning what I learn into an application. My interests connect machine learning, explainable AI, computer vision, and the backend systems that make models usable.
I'm eager to learn from others, work through difficult problems, and contribute to a team. My project reports show the questions I investigate, the approaches I compare, and the limitations I take seriously.
- Understand the problem: identify the data, failure cases, and practical cost of getting a prediction wrong.
- Test the alternatives: compare baselines and architectures under consistent evaluation conditions.
- Look beyond accuracy: inspect per-class recall, confusion matrices, imbalance, and the trade-off between performance and efficiency.
- Explain the result: explore SHAP, LIME, and Grad-CAM to investigate what influences predictions.
- Build something usable: connect experiments to Flask or FastAPI interfaces, then document setup, evidence, and limitations.
| Project | Problem & approach | Stack |
|---|---|---|
| Reliable 4G/5G Radio Resource Forecasting Under Missing Telemetry | Real 4G/5G resource forecasting: chronological evaluation, calibrated ranges, missing-input checks and a historical replay dashboard | scikit-learn · LightGBM · SQLite · JavaScript |
| V4FinBench | Financial distress under class imbalance: grouped validation, temporal testing, and feature-timing audits | LightGBM · XGBoost · CatBoost · SHAP |
| Heavy Vehicle Predictive Maintenance | Rare failure classification: synthetic telemetry, ensemble comparisons, cost-based evaluation, and a Flask interface | scikit-learn · SMOTE · LIME · Flask |
| Brain Tumor Classification | MRI classification reliability: compare CNNs, investigate class-specific errors, and explore ensemble explanations | TensorFlow · Keras · Grad-CAM · Flask |
| IncidentLab | Investigating cloud incidents with multimodal telemetry and a model-serving API | PyTorch · FastAPI · Docker · Kubernetes |
| CampusCart | Equipment reservations through clear APIs and containerized microservices | FastAPI · PostgreSQL · Docker · Kubernetes |
CNN architecture comparison: I implemented and trained InceptionV3 and MobileNetV3. Naga Sreerama Pradyumna Tata implemented EfficientNetB3 and DenseNet201. We jointly handled preprocessing, performance evaluation, confusion matrix analysis, and report writing. See the deep-learning report.
Explainable brain-tumor classification: coauthor of the ensemble and reliability study with Yashwanth Hudumula, and a named team member in the major-project report.
Predictive maintenance: coauthor with Naga Sreerama Pradyumna Tata of the heavy-vehicle report, covering preprocessing, imbalance, model comparisons, evaluation, and explainability.
These are academic reports. Results belong to their stated experiments; repository READMEs explain the saved implementation evidence and limitations.
Intern — Nukkad Shops Technologies (India) Private Limited
May 13, 2025 – August 12, 2025 · Completion certificate
Why does a model fail on one class while its overall accuracy looks strong? What happens when the data shifts? How can explanations and APIs help people understand and use a system?
Those questions keep me experimenting, learning, and building.