Surya
Prabhav
Gurram

iOS application that converts Apple Watch ECG recordings into privacy-preserving, on-device stress estimates and streams feature-level telemetry over MQTT for IoT scenarios. Implements a full Edge ML pipeline: ECG-to-RR extraction via HealthKit, robust RR preprocessing with outlier handling, time-domain HRV feature extraction (RMSSD, SDNN, pNN50, SD1/SD2), and two stress estimators — an interpretable heuristic score and a personalized logistic regression model with isotonic calibration trained directly on the user's own tagged sessions. No raw biometric data leaves the device.

An intelligent index recommendation framework that addresses the notoriously hard problem of index selection in relational databases. Uses workload-driven analysis on industry-standard TPC-H benchmark datasets to profile query patterns and access frequencies. Implements automated benchmarking pipelines to measure before/after query latency with precision. Draws on BALANCE-inspired multi-objective analysis and λ-Tune optimization concepts to balance index benefit against storage and write overhead — producing recommendations that mirror what a senior DBA would manually derive.

Deep learning system for automated brain tumor diagnosis across two imaging modalities. The MRI pipeline uses CNN-based models for tumor detection and precise scan segmentation, isolating regions of interest with high spatial accuracy. A parallel hyperspectral imaging pipeline performs tissue-level spectral analysis, differentiating healthy from malignant tissue based on wavelength signatures invisible to standard imaging. Incorporates multi-stage image preprocessing — noise reduction, normalization, and augmentation — to improve generalization on limited medical datasets.

End-to-end smart farming platform that applies machine learning across the agricultural lifecycle. Integrates multi-variable crop yield prediction using ensemble models trained on soil, climate, and historical yield data; real-time weather forecasting via API integration; and a CNN-based plant disease detection system that classifies leaf images into 38 disease categories with high accuracy. The Django/Flask backend serves predictions to a responsive frontend, delivering actionable farming recommendations to reduce crop loss and optimize resource usage.

Probabilistic fraud detection system combining two complementary Bayesian approaches: a Hybrid Naive Bayes classifier for fast, scalable transaction scoring, and a Bayesian Belief Network that models causal dependencies between transaction features to capture complex fraud patterns. The preprocessing pipeline handles severe class imbalance via stratified sampling and feature engineering on transactional sequences. Evaluated against standard baselines, the hybrid approach demonstrates improved precision-recall trade-offs critical for minimizing both false positives (declined legitimate transactions) and false negatives (missed fraud).

CNN-based diagnostic system for cardiovascular disease screening from ECG imagery. Rather than relying solely on raw waveform amplitude, the model extracts features across both temporal and frequency domains — capturing rhythm irregularities, QRS complex morphology, and spectral signatures associated with specific cardiac conditions. The pipeline processes 12-lead ECG images through a multi-scale convolutional architecture, enabling the model to learn both local waveform features and global rhythm patterns simultaneously. Designed for deployment in resource-constrained clinical settings.