Method: Built a FastAPI + Streamlit workflow that gathers public-source evidence, resolves subject ambiguity, classifies evidence against a fixed rubric, and computes final risk outcomes with deterministic rules.

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

Method: Uses local résumé parsing/OCR, explicit consent before external structured extraction, editable human review, a checksum-validated DistilBERT seed-42 classifier for broad occupation recognition, and a separate deterministic target-role evidence layer.

Output: A reviewable profile, occupation suggestions, supported and missing target-role evidence, plus transition tools; the classifier does not choose the user's target role or claim hiring probability.

Method: Uses a clinical knowledge graph and explicit rules for decisions, RAG for grounded evidence, and an LLM only for language understanding and explanation; my contribution focused on the RAG pipeline, retrieval evaluation, grounded prompts, output validation, and hallucination control.

Output: Personalized, traceable rehabilitation guidance with multi-layer evaluation across decision classification, rehabilitation phase, recommendation quality, and user-condition scenarios.

Method: Aligned the target to guest count, removed leakage-prone fields, made production retraining require explicitly staged real data, selected models on validation RMSE against a mean-baseline gate, locked holdout testing, and fingerprinted promoted artifacts.

Output: Immutable run/model artifacts with source lineage and SHA-256 verification, drift and prediction-shift monitoring, authenticated FastAPI online/batch inference, structured request logs, Prometheus metrics, non-root Docker, Kubernetes probes/HPA/PDB, and documented rollback controls.

Method: Derived a leakage-safe Fail/Pass target from Failure A-E, compared Logistic Regression, Random Forest, and Gradient Boosting, tuned the operating threshold on validation F2, reserved the test split for final evaluation, and fingerprinted candidate model artifacts for immutable promotion.

Output: A versioned preprocessing/model/threshold bundle plus deployment manifest, SHA-256 verification, authenticated bounded-batch FastAPI scoring, structured request logs and IDs, Prometheus metrics, readiness/liveness probes, non-root Docker, Kubernetes HPA/PDB, CI image builds, and rollback guidance; no new performance number is claimed until the source dataset is rerun.

Method: Used stratified train/validation/test splits, compared Logistic Regression, Random Forest, and LightGBM by validation AUC, tuned the operating threshold only on validation data, tracked experiments with MLflow, and fingerprinted the promoted preprocessing/model/threshold bundle.

Output: A versioned LightGBM serving bundle with locked-test AUC 0.9253, SHA-256 verification, authenticated bounded-batch FastAPI inference, structured request logs and IDs, Prometheus metrics, data-quality/drift/business monitoring, non-root Docker, Kubernetes HPA/PDB, CI image builds, and rollback guidance.

Method: Cleaned invalid and missing inputs, built numeric/categorical preprocessing, compared Logistic Regression, Random Forest, and Extra Trees on stratified train/validation/test splits, selected the operating threshold on validation phishing F2, and reserved the test split for locked evaluation.

Output: A versioned model bundle with feature schema and training reference profile, SHA-256 deployment manifest, authenticated single/batch FastAPI inference, numeric PSI and categorical drift checks, structured request logging, Prometheus metrics, non-root Docker, Kubernetes probes/HPA/PDB, CI container builds, and rollback guidance; no performance metric is claimed until the intended dataset is rerun.