Nikita Gajbhiye Nikita3005

_{DATA SYSTEMS / APPLIED ML / TRUSTWORTHY AI}

Nikita Gajbhiye

Data Science · Machine Learning · Data Engineering

I build reliable data and ML systems, with a focus on distribution shift, rigorous evaluation, and secure AI agents.

Featured projects

🛡️ TaintGate

A runtime guard that evaluates protected AI agent tool calls before execution.

Method: Tracks input provenance, evaluates security findings, and applies deterministic policy to return ALLOW, REVIEW, or BLOCK.

Validation: A local regression suite checks expected decisions across 12 included attack scenarios. The repository also includes framework and MCP integrations.

Code and attack suite →

📉 DriftForge

A benchmark investigating early warnings of synthetic-data-induced model degradation.

Design: 3 datasets × 3 mechanisms × 5 seeds × 11 levels = 495 controlled conditions.

Evaluation: Compares drift signals using discrimination and warning lead time, and examines whether thresholds transfer across datasets.

Code → · Technical report →

🧩 Enterprise Customer Intelligence

A modular customer ML platform covering churn prediction, lifetime value, and segmentation.

System: Feature engineering and model evaluation connect to SHAP explanations, MLflow experiment tracking, and FastAPI serving, with automated tests and CI.

Models and system design →

🏗️ Supply Chain Risk Intelligence

A shipment data platform for operational risk analysis.

Architecture: Databricks Bronze, Silver, and Gold layers prepare shipment data for delay prediction, vendor risk analysis, and SQL dashboards.

Pipeline and analytics →

Additional work

Project	Technical focus
HealthLynked Provider Pipeline	Record changes · Duplicate detection · Confidence scoring · Human review
Fraud Detection & Risk Modeling	Imbalanced classification · Logistic regression · Random forests · Precision–recall
Customer Churn Analytics	Customer behavior · Churn prediction · Power BI reporting

Technical snapshot


495 controlled conditions _{DriftForge ↗}	12/12 local attack checks _{TaintGate ↗}	284,807 transactions in dataset _{Fraud Detection ↗}