Parusha Sindhu

Parusha Sindhu

Machine Learning Engineer focused on building interpretable and fair AI systems

I design ML models that don’t just predict — they explain decisions, uncover bias, and ensure reliable outcomes in real-world applications.

97%

Model Accuracy

Production Projects

3 Years

NCC Cadet

Why Explainability & Fairness Matter

In production ML systems, a model's accuracy is only half the story. Stakeholders need to understand why a decision was made, and regulators demand proof that systems don't discriminate.

I specialize in building ML pipelines that are both high-performing and interpretable — using SHAP and LIME to surface feature importance, and fairness frameworks to audit bias before deployment.

Production-Ready Code

Clean, documented pipelines that integrate with existing systems

Model Interpretability

SHAP and LIME explanations that non-technical stakeholders can understand

Fairness Auditing

Quantified bias metrics across demographic groups pre-deployment

XAI Cultural Heritage Prediction

Explainable AI Multi-class Classification SHAP

Developed an ML pipeline to predict cultural heritage site significance using ensemble models (XGBoost, CatBoost). Integrated SHAP for per-prediction feature attribution, enabling archaeologists and policy-makers to understand why each site received its classification. Delivered interactive visualizations showing how region, material composition, and structural age influence predictions.

SHAP-driven explanations for every prediction

Ensemble models for robust classification

Interactive dashboards for stakeholders

View on GitHub

Fairness AI Dashboard

Full-Stack Bias Auditing MongoDB

Built a production-grade bias evaluation system to audit trained models across demographic groups. Computes disparate impact, equalized odds, and demographic parity metrics in real-time. Backend stores audit logs in MongoDB; frontend dashboard visualizes fairness gaps and flags models that fail regulatory thresholds before deployment.

Real-time bias metric computation

MongoDB audit trail & versioning

Group-level performance comparison

View on GitHub

Fetal Health Classification

Healthcare ML CatBoost 97% Accuracy

Trained a high-precision CatBoost classifier on cardiotocography data to detect fetal distress across three health classes (Normal, Suspect, Pathological). Achieved 97.2% accuracy through feature engineering, SMOTE-based class balancing, and Bayesian hyperparameter tuning. Validated with confusion matrix analysis showing >91% recall on critical Pathological class.

97.2% accuracy, 91%+ recall on critical class

CatBoost with Bayesian optimization

Full confusion matrix + precision/recall

View on GitHub

Why Explainability & Fairness Matter

Production-Ready Code

Model Interpretability

Fairness Auditing

Key Projects

XAI Cultural Heritage Prediction

Fairness AI Dashboard

Fetal Health Classification

Technical Expertise

Machine Learning

Explainability & Fairness

Data & Infrastructure

Visualization

Leadership & Discipline

Let's Work Together