ARCeH logoARCeH Externship via Extern

Data analytics for health outcomes, in progress

I am a professional specializing in Data, AI, and Machine Learning, focused on delivering innovative solutions that drive efficiency and enhance decision-making.

GitHubMessage Me

Work samples

Portfolio of projects completed during my ARCeH externship and related graduate work, focused on data analysis, health outcomes questions, and practical reporting.

  • Final Capstone Project Baseball(A)IQ APP
    Final Capstone Project Baseball(A)IQ APP

    Designed and built a regression pipeline predicting MLB players' next-season batting average from prior performance, engineering a player-season panel dataset from the Lahman Baseball Database (1962–2023) with a time-aware train/test split to prevent data leakage Engineered a sabermetrically-grounded feature set (BABIP, walk/strikeout rate, age curve, positional context) to capture predictive signal beyond raw batting average alone Trained and compared 5 model configurations (Linear Regression, Random Forest, Gradient Boosting) across algorithm and feature-set variations, tracking experiments with MLflow and selecting the best model by MAE/R² [Tuned Linear Regression- MAE: 0.0234, RMSE: 0.0298, R2: 0.2247] Built a natural-language interface (Streamlit + LLM API) supporting both real-player lookups and hypothetical "what-if" scenarios, with structured input parsing, clarifying questions for missing data, and graceful handling of out-of-scope queries Applied production MLOps practices: config-driven training (YAML), DVC-tracked datasets, a pytest suite covering preprocessing, model validation, and interface logic

    MLflowStreamlitLLM APIYAMLDVCpytest
  • End to End MLops Pipeline
    End to End MLops Pipeline

    Designed and built an end-to-end MLOps pipeline for a healthcare classification model Git, DVC, MLflow, pytest, GitHub Actions, Evidently — covering raw data ingestion through CI/CD-gated training Implemented DVC-based dataset versioning with an in-repository remote, enabling reproducible data access (dvc pull) from any clone with no external cloud dependencies Tracked and compared 5+ ML experiments across two model types in MLflow, with automated best-run selection via mlflow.search_runs() Built a 14-test pytest suite spanning unit, data validation, and model validation tiers, achieving a 100% pass rate under CI Configured a two-job GitHub Actions CI/CD pipeline (test → train) that automatically fails builds when model performance drops below defined thresholds Implemented Evidently-based drift detection comparing simulated production data against training data, generating HTML reports and enforcing a configurable drift-share threshold Root-caused a cross-platform data corruption bug where CRLF line-ending conversion silently broke DVC's content-addressed storage on Linux CI runner Diagnosed a Python import-path defect that only surfaced under the project's exact required test-invocation command

    GitDVCMLflowpytestGitHub ActionsEvidently

About me

I am a professional specializing in Data, AI, and Machine Learning, focused on delivering innovative solutions that drive efficiency and enhance decision-making.

I am AJ Bova, a Data professional and aspiring AI/ML Engineer with practical experience in Python, statistical analysis, and machine learning. With over 10 years in sales management and business development, I excel at turning complex data into actionable insights. I have a strong track record of driving revenue growth, negotiating high-value contracts, and leading effective teams. I am committed to using data-driven decision-making to foster positive change for organizations moving towards AI-powered operations.

Externships

Data Analytics, Health Outcomes Externship with ARCeH

ARCeH

Experience

Outside Sales Representative

Jacquet Mid Atlantic · 06/2024 - Present

OEM Sales Manager

Neoperl Inc. · 02/2023 - 06/2024

Inside Sales Manager

Thyssenkrupp Materials NA · 05/2017 - 02/2023

Education

TripleTen

AI & Machine Learning Bootcamp · Class of 2026

Southern New Hampshire University / Franklin Pierce University

Business Management & Business Administration

Skills

Data analysisHealth outcomes researchGraduate-level study in analyticsExternship project workBusiness DevelopmentBusiness StrategyBusiness IntelligenceCI/CDClaudeCoachingCollaborationCommunicationCompany AnalysisCompetitive AnalysisComputer VisionCRMData AnalyticsData VisualizationData ScienceExcel
Back to work

Final Capstone Project Baseball(A)IQ APP

Designed and built a regression pipeline predicting MLB players' next-season batting average from prior performance, engineering a player-season panel dataset from the Lahman Baseball Database (1962–2023) with a time-aware train/test split to prevent data leakage Engineered a sabermetrically-grounded feature set (BABIP, walk/strikeout rate, age curve, positional context) to capture predictive signal beyond raw batting average alone Trained and compared 5 model configurations (Linear Regression, Random Forest, Gradient Boosting) across algorithm and feature-set variations, tracking experiments with MLflow and selecting the best model by MAE/R² [Tuned Linear Regression- MAE: 0.0234, RMSE: 0.0298, R2: 0.2247] Built a natural-language interface (Streamlit + LLM API) supporting both real-player lookups and hypothetical "what-if" scenarios, with structured input parsing, clarifying questions for missing data, and graceful handling of out-of-scope queries Applied production MLOps practices: config-driven training (YAML), DVC-tracked datasets, a pytest suite covering preprocessing, model validation, and interface logic

MLflowStreamlitLLM APIYAMLDVCpytest
View all work
Back to work

End to End MLops Pipeline

Designed and built an end-to-end MLOps pipeline for a healthcare classification model Git, DVC, MLflow, pytest, GitHub Actions, Evidently — covering raw data ingestion through CI/CD-gated training Implemented DVC-based dataset versioning with an in-repository remote, enabling reproducible data access (dvc pull) from any clone with no external cloud dependencies Tracked and compared 5+ ML experiments across two model types in MLflow, with automated best-run selection via mlflow.search_runs() Built a 14-test pytest suite spanning unit, data validation, and model validation tiers, achieving a 100% pass rate under CI Configured a two-job GitHub Actions CI/CD pipeline (test → train) that automatically fails builds when model performance drops below defined thresholds Implemented Evidently-based drift detection comparing simulated production data against training data, generating HTML reports and enforcing a configurable drift-share threshold Root-caused a cross-platform data corruption bug where CRLF line-ending conversion silently broke DVC's content-addressed storage on Linux CI runner Diagnosed a Python import-path defect that only surfaced under the project's exact required test-invocation command

GitDVCMLflowpytestGitHub ActionsEvidently
View all work
Back to work

RAG-Powered Knowledge Assistant

Built an end-to-end Retrieval-Augmented Generation pipeline from scratch — document chunking, vector embeddings, semantic search, and extractive question-answering — using an 8-document, 2,100+ word knowledge base processed into 53 sentence-aware chunks with configurable overlap Implemented semantic retrieval with ChromaDB and Sentence-Transformer embeddings (all-MiniLM-L6-v2, 384-dim), combined with a confidence-gated response layer to distinguish answerable from out-of-scope queries Diagnosed and fixed a silent failure mode where out-of-scope questions returned high-confidence fabricated answers; resolved by adding an embedding-distance relevance gate, calibrated against real retrieval data, reducing false-positive confident answers from 100% to 0% across edge-case testing Debugged a transformers/PyTorch version incompatibility by replacing the high-level QA pipeline with a direct model inference implementation (manual logit extraction and span decoding), maintaining full functionality without downgrading dependencies Evaluated system performance across 15 test queries in 4 categories, identifying and documenting a structural limitation of extractive QA on comparison/multi-part questions and proposing a generative-model upgrade path

ChromaDBSentence-TransformertransformersPyTorch
View all work
Back to work

Medical Insurance Prediction Neural Network

Built and trained a deep learning regression model in Keras to predict individual medical insurance costs from demographic and lifestyle features (age, BMI, smoker status, region, etc.) Conducted exploratory data analysis that uncovered a significant smoker-status X BMI interaction driving cost outliers, informing feature preprocessing and model design decisions Diagnosed a systematic model weakness traced to function behavior on skewed target data resolved it via log-transformation, improving overall prediction accuracy by 13% and reducing relative error imbalance across the full cost range. Validated through controlled experimentation across multiple architectures to confirm optimal model complexity for the data size. Applied full ML best practices: train/test splitting prior to normalization to present data leakage, one-hot encoding of categorical features, and iterative model evaluation using training/validation loss curves and residual error analysis.

Keras
View all work
Back to work

House Price Prediction API-FastAPI Deployment Service

Designed and deployed a production-style REST API to serve a trained machine learning model (Linear Regression, 13 features) using FastAPI, Pydantic, and Uvicorn Implemented model persistence and versioning (joblib serialization, semantic versioning, metadata tracking) to support reproducible model updates Built input validation and structured error handling using Pydantic schemas, ensuring reliable client-server data contracts Engineered efficient model loading (single load at startup vs. per-request) to optimize response latency Delivered self-documenting API endpoints (/predict, /model/info, /health) with interactive Swagger UI documentation

FastAPIPydanticUvicornjoblib
View all work
Back to work

Beta Bank Customer Turnover Pediction

Built a binary classification model to predict customer churn from 10,000 customer records, addressing significant class imbalance (80/20 split) using scikit-learn Engineered and compared class-imbalance correction techniques — downsampling, upsampling, and class_weight='balanced' — across Logistic Regression, Decision Tree, and Random Forest models Selected and tuned a final Random Forest model (n_estimators=100, max_depth=10, class_weight='balanced', threshold=0.47), achieving F1 score of 0.607 and AUC-ROC of 0.854 on the test set Applied feature engineering techniques (interaction/ratio features, StandardScaler) and threshold tuning to optimize model performance beyond the required F1 ≥ 0.59 benchmark Refined methodology through iterative code review, incorporating additional baseline models and resampling techniques per reviewer feedback

scikit-learn
View all work
Back to work

Megaline Mobile Plan Classification

Built a classification model to recommend optimal mobile plans ("Smart" vs. "Ultra") for 3,000+ customers based on usage behavior, using scikit-learn Evaluated three algorithms — Decision Tree, Random Forest, and Logistic Regression — selecting Random Forest (n_estimators=200, max_depth=7) as the final model, achieving 81.6% accuracy Implemented stratified train/validation/test splitting (60/20/20) to preserve class balance and prevent data leakage during hyperparameter tuning Refined model development through iterative code review, correcting methodology to ensure the test set was evaluated only once for unbiased final performance reporting

scikit-learn
View all work
Back to work

Venture Capital Investment Analysis

Performed end-to-end SQL analysis on a comprehensive venture capital database for VentureInsight, a research firm serving VC clients making multi-million dollar investment decisions. Queried and analyzed a 7-table relational database spanning companies, funds, funding rounds, investments, acquisitions, and people to deliver insights for a quarterly investment report Engineered complex multi-table queries using INNER JOINs, subqueries, and CTEs to analyze employee education levels at failed startups and identify correlations between team background and company outcomes Built fund activity classification system using CASE logic to categorize venture funds into high, middle, and low activity tiers enabling clients to identify appropriate co-investment partners Conducted geographic and sector funding analysis using GROUP BY aggregations to identify top funded countries and US news sector investment benchmarks for international investment strategy decisions

SQL
View all work
Back to work

Video Game Sales Forecasting

Analyzed historical video game sales data for a simulated online retailer to identify patterns determining game success and inform a 2017 advertising campaign strategy. Engineered a data pipeline in Python using Pandas to clean, preprocess, and transform a 16,700+ record dataset including handling missing values, type conversions, and TBD entries. Conducted platform and genre sales analysis across 12 platforms and 12 genres to identify PS4 and Xbox One as the highest priority platforms for budget allocation. Built regional user profiles for North America, Europe, and Japan using grouped aggregations to reveal distinct genre preferences by market. Executed independent samples t-tests to validate hypotheses about user rating differences across platforms and genres, providing statistically grounded marketing recommendations.

PythonPandasMatplotlib
View all work
Back to work

Statistical Data Analysis of Megaline Consumer Data

Analyzed usage and revenue data from 500 Megaline telecom clients to determine which prepaid plan generates more revenue and guide advertising budget decisions. Preprocessed and aggregated multi-table customer data in Python including rounding rules, monthly usage calculations for calls, texts, and data consumption. Computed statistical metrics including mean, variance, and standard deviation to compare revenue performance across Surf and Ultimate prepaid plans. Conducted hypothesis testing on revenue differences by plan and by region to deliver statistically supported advertising recommendations. Visualized key trends using Matplotlib to clearly communicate findings to non-technical stakeholders.

PythonMatplotlib
View all work
Back to work

Exploratory Data Analysis — Instacart Customer Behavior

Performed end-to-end EDA on Instacart's grocery orders dataset to uncover shopping patterns, peak ordering times, reorder behavior, and top products. Cleaned and preprocessed multiple relational tables in Pandas addressing missing values, duplicate entries, and data type inconsistencies. Identified peak ordering windows and top reordered products through grouped aggregations and frequency analysis. Designed and produced a suite of Matplotlib visualizations to communicate customer behavior insights clearly and effectively.

PandasMatplotlib
View all work