Vishwa Hatwalne

Created Duet, a 185,881-sample Windows-malware dataset spanning six years and eight behaviour categories, pairing static features with CAPEv2-derived dynamic behaviour graphs. Developed MalCross - a field-wise static Transformer, Dynamic Edge-Conditioned Encoder and modality-preserving cross-interaction - with reproducible training, ablation and benchmark pipelines. Achieved 67.46 macro-F1, 12.56 points above the strongest baseline, and introduced the Ordinal Discordance Index.

Developed reusable PyTorch training, validation, inference and persistence components for a three-layer BiLSTM-attention State-of-Charge estimator. Evaluated an offline TD3+BC uncertainty-adaptation agent, achieving about 1.446% held-out SOC MAE in the Q-only experiment. Built leakage-safe leave-one-battery-family-out evaluation and automated experiment pipelines. Processed about 146 million telemetry observations and created over 1,000 physical state-anchor events.

Developed a real-time Python pipeline for streaming respiratory waveforms, detecting peaks, troughs and trigger points with adaptive thresholds, state transitions, cycle constraints and defensive handling of incomplete data. Designed evaluation using hit rate, sample-level timing error, matching windows, per-event diagnostics and descriptive statistics. Built incremental PyQtGraph/Qt and Plotly diagnostics plus reusable pandas/NumPy ingestion and export pipelines.

Fine-tuned multilingual automatic speech recognition models with Kaldi C++; reduced Word Error Rate by 0.3 points through language-model tuning and feature extraction. Used decoding graphs and phoneme alignments for systematic error analysis.

Engineered maintainable Python backend components for a high-priority global warehouse-management platform. Built automated test suites with pytest, unittest and mock to improve CI reliability; contributed to Airflow and MLflow workflows for model versioning, monitoring and deployment.

Developed CatBoost fraudulent-shipment models and partnered with data engineers and product managers to define KPIs and productionise outputs. Built an ETL pipeline that improved downstream model accuracy by 15%; vectorised legacy pandas/NumPy processing to reduce compute time by 70%.

Data Science M.Sc. candidate; expected completion October 2026.

Bachelor of Technology in Computer Science & Engineering.