Siddhanta Adhikari

Curated a custom dataset of Nepali folk music audio samples for regional genre classification, with no existing benchmark dataset available. Built a hybrid EfficientNet-B0 + BiLSTM classifier for regional genre identification, processing raw audio into mel spectrogram representations and benchmarking against an SVM baseline.

Directed development of a UNet++ semantic segmentation model achieving macro IoU 0.734 and macro F1 0.829 across 11 flood events, later formalized into a manuscript submitted to the Journal of Hydroinformatics. Benchmarked 4 architectures and built a validation pipeline integrating ESA WorldCover permanent-water masks.

Designed a normalized PostgreSQL schema and ER diagrams to support multi-user recipe management, exposed through a Django REST backend with full CRUD API endpoints. Built a responsive React frontend connecting relational data models end-to-end from database to UI.

Built an ETL pipeline in Python that pulled 188,000+ historical F1 records (2018–2025) from two APIs, cleaned and standardized the data, and validated data quality. Ran exploratory data analysis on race results, pit stops, and tire strategy to uncover patterns and produce charts.

Engineered a security middleware layer that intercepts and approves every AI tool call before execution, addressing OWASP LLM-07 and LLM-08 risks. Integrated Llama 3.2 via Ollama, running tool execution in an isolated, network-disabled Docker container with validation checks and adversarial tests.

Grade: 3.913

Pursuing a Bachelor of Engineering in Computer Engineering with a GPA of 3.913, expected to graduate in February 2027.