Data Science · AI · Software Engineering
I build across machine learning, generative AI, data engineering, automation, backend systems, and application development. My work focuses on complete technical systems: problem formulation, data architecture, implementation, validation, testing, limitations, and explainability.
- AI & Machine Learning: TensorFlow, Keras, CNNs, transformers, Hugging Face, NLP, Sentence-BERT, HDBSCAN, interpretable ML, calibration, subgroup validation
- Data Engineering & Analytics: Python, Pandas, ETL, semi-structured data processing, feature engineering, statistics, anomaly detection, Power BI, Power Query, DAX, public-health analytics
- Backend & Databases: Flask, REST APIs, SQL, PostgreSQL, SQLite, MySQL, SQLAlchemy, Alembic, relational modeling, transactions, RBAC, concurrency controls, testing
- Application Engineering: TypeScript, Docker, Docker Compose, service integration, browser/client runtime boundaries, PDF processing, financial analytics, adaptive recommendation systems
Modern reconstruction of a recovered CSIT-558 Data Mining project using CDC/NHANES public-use data. Builds a reproducible adult blood-pressure analytics pipeline, an interpretable logistic-regression benchmark, calibration and subgroup validation, and governed random-forest/gradient-boosting challengers.
Tech: Python · Pandas · scikit-learn · Logistic Regression · Model Validation · Calibration · Model Governance
Provenance-aware document-intelligence project for public U.S. presidential documents. The public implementation combines deterministic priority rules with SentenceTransformer semantic similarity, a 14-domain taxonomy, primary/secondary/tertiary domain signals, review and source-quality heuristics, isolated HTML/PDF retrieval helpers, and Power BI-oriented outputs. Recovered historical V1–V3 source lineage is preserved through explicit provenance documentation while private work-specific paths remain excluded.
Tech: Python · Pandas · SentenceTransformers · all-MiniLM-L6-v2 · BeautifulSoup · pypdf · Power BI · DAX · Power Query · pytest · GitHub Actions
Personal-finance analytics application with browser-side CSV and searchable-PDF ingestion, transaction normalization, robust spending baselines and deviation screening, financial-goal/investment scenario modeling, and evidence-calibrated recommendation signals. The enhanced implementation includes deterministic dependency locking, dependency auditing, expanded tests, and Node.js 24 CI.
Tech: TypeScript · Vite · PDF.js · Financial Analytics · Robust Statistics · Adaptive Recommendations · Vitest
Sanitized Python ITS/SCATS operator-support workflow for deterministic Green-Time screening against row-level baselines, configurable severity thresholds, operational input-quality metrics, daily batch reporting, and auditable Excel outputs. The public implementation preserves the conceptual separation between Green-Time screening and phase diagnostics without claiming unavailable production multi-export integrations.
Tech: Python · Pandas · openpyxl · ITS/SCATS · Operational Analytics · Data Quality · pytest
Reconstructed and enhanced Amazon Reviews 2023 NLP pipeline using FLAN-T5 for prompt-conditioned cleaning, one-sentence summarization, and feature/issue extraction, plus a BERT-based 1–5 star sentiment model. The implementation adds token-aware truncation, structured CSV/JSON outputs, ordinal evaluation, deterministic offline tests, dependency auditing, and multi-version CI.
Tech: Python · Hugging Face · FLAN-T5 · BERT · NLP · scikit-learn · pytest
Reconstructed and enhanced behavioral-analytics pipeline that converts semi-structured Excel activity diaries into auditable event-level and participant-day data, clusters behavioral routine features with HDBSCAN, and supports TF-IDF, optional Sentence-BERT weekday/weekend comparison, deterministic t-SNE visualization, data-quality diagnostics, and descriptive mood-association analysis. Mood is deliberately excluded from clustering inputs to preserve interpretability.
Tech: Python · Pandas · HDBSCAN · Sentence-BERT · TF-IDF · t-SNE · SciPy · pytest
Transparent TensorFlow/Keras reconstruction of a recovered 20,000-image cat-vs-dog CNN experiment. The historical run logged 97.23% training accuracy and 82.30% validation accuracy, exposing substantial overfitting. The enhanced repository adds memory-efficient tf.data ingestion, image-integrity checks, precision/recall/F1/ROC-AUC diagnostics, structured run artifacts, 25 automated tests, Ruff, dependency auditing, and a TensorFlow runtime CI gate.
Tech: TensorFlow · Keras · CNN · Computer Vision · scikit-learn · pytest
Reconstructed and substantially enhanced Flask/SQLAlchemy registration backend with terms, sections, prerequisites, schedule validation, FIFO waitlists, RBAC, Alembic migrations, PostgreSQL final-seat row locking, real concurrency integration testing, audit logging, rate limiting, Prometheus metrics, OpenAPI/Swagger documentation, and Docker/Gunicorn deployment support.
Tech: Flask · SQLAlchemy · PostgreSQL · Alembic · REST APIs · RBAC · Transactions · Docker · pytest
Verified CSIT 555 Database Systems team project from Spring 2024. The original six-person project used Flask, MySQL, Docker Compose, CRUD workflows, reporting, and service integration. The current portfolio repository clearly separates that recovered history from a focused Flask/SQLite reconstruction for user-entered quarterly obligations and partial payments, with integer-cent money handling, relational constraints, derived payment status, focused tests, and CI.
Tech: Current — Flask · SQLite · Decimal/integer-cent money handling · pytest · GitHub Actions · Historical — MySQL · Docker Compose · CRUD/reporting workflows
I emphasize evidence over embellishment. My portfolio distinguishes:
- RECOVERED — supported by prior work, records, or artifacts
- RECONSTRUCTED — faithfully rebuilt when original source bytes were unavailable
- ENHANCED — modern engineering improvements added for portfolio quality
- UNVERIFIED — not presented as historical fact without supporting evidence
This keeps technical claims accurate while still demonstrating full current implementation ability.