Problem: Public-source reputational screening can be slow, inconsistent, and difficult to audit.

Method: Built a FastAPI + Streamlit workflow that gathers evidence, applies a fixed risk rubric, and uses deterministic scoring rules.

Output: Reviewer-ready findings, supporting evidence, risk assessment, and memo-style reporting for more consistent compliance decisions.

Problem: Mid-career users need to separate broad occupation similarity from evidence that actually supports a specific target role.

Method: Combined résumé extraction/OCR with a 12,000-feature TF-IDF + balanced Logistic Regression classifier and separate deterministic target-role evidence matching.

Output: A reviewable profile, occupation suggestions, supported and missing target-role evidence, plus job-search, interview, and transition-planning tools.

Problem: Home-based knee rehabilitation can lack continuous support, increasing the risk of unsafe or inappropriate recovery choices.

Method: Used a clinical knowledge graph and explicit rules for decisions, RAG for grounded evidence, and the LLM only for language understanding and explanation.

Output: Personalized, traceable rehabilitation guidance with evidence-grounded explanations and scenario-based safety validation.

Problem: Demand signals were split across search and contact interactions, requiring a reproducible demand proxy and model lifecycle.

Method: Joined and validated source data, benchmarked a mean baseline against Linear Regression and XGBoost, then packaged the selected pipeline with monitoring.

Output: Holdout metrics, explainability artifacts, a versioned model bundle, run manifests, drift reports, alerts, and retraining/rollback assets.

Problem: Missing and non-physical fleet data made vehicle failure risk difficult to interpret reliably.

Method: Used DuckDB SQL, Python, and pandas to clean/impute operating fields, engineer a Fail/Pass target and risk bands, and compare segment failure rates.

Output: A cleaned ML-ready dataset and preventive-maintenance insights across vehicle, factory, usage, membership, and operating-condition segments; no trained classifier is claimed.

Problem: Imbalanced browsing sessions make purchase intent difficult to detect while balancing missed purchasers against false positives.

Method: Built stratified train/validation/test preprocessing, compared Logistic Regression, Random Forest, and LightGBM by validation AUC, tuned the threshold, and tracked experiments with MLflow.

Output: A LightGBM final model with held-out AUC 0.9253, threshold analysis, feature importance, FastAPI/Docker serving, and lightweight drift, data-quality, and business-impact monitoring.