Projects

A Windows desktop app for exploring large data files. Load CSV, TSV or JSON into one SQLite dataset, then query and chart it by pointing and clicking.

Problem

Data-competition teams get large raw files, and not everyone on the team knows SQL.

Approach

Load every file into one SQLite dataset and generate the SQL from point-and-click choices, so teammates can explore the data without writing queries. Generated queries are read-only, and accounts and roles guard the data.

Implementation

• A streaming importer for CSV, TSV, JSON and JSON Lines, with column type detection and a two-thread parse and write pipeline.
• A relationship finder that infers foreign keys from column names and matching values, rates each link’s confidence and draws the links as a diagram.
• A query builder that turns point-and-click joins, filters and sorting into read-only SELECT statements, with bar, line, pie and scatter charts.
• Accounts and roles. The first launch creates the owner. Passwords are hashed with Argon2id and need 8+ characters, a digit and a symbol. There are no default passwords. Roles run USER, ADMIN, OWNER.
• 182 JUnit tests. JavaFX written in code rather than FXML. Shipped as a Windows app through GitHub Releases.

Result

A complete Windows app shipped through GitHub Releases. It imports an 8-million-row, 1.3 GB CSV in about 3.5 minutes, with 182 JUnit tests behind it.

A C++20 console tool that cleans large CSV files. The multithreaded rewrite handles a 1.47 GB file in 2.2 s. The original took 447.5 s as a Debug build.

Problem

The competition’s 1.47 GB patient-encounter CSV (7,675,802 lines) was too large to open in Excel or Access.

Approach

Drop every row with a placeholder value so the file is small enough to analyze. The original was mine and single-threaded. I did the rewrite and its review with Claude Code: correct first, measure each change, keep output byte-identical.

Implementation

• Cleans every .csv in a folder. It drops any row with a placeholder value (*deleted, *not applicable, *unspecified, *unknown, NA), adds an original_row_index column and writes name_cleaned.csv next to the input. It handles commas inside quotes and Windows line endings.
• The DataFest version was single-threaded. The rewrite uses one reader thread that splits the file into 4 MB blocks ending on a line break, worker threads (cores minus 2, so 12 on the test machine) that clean blocks in parallel, and the main thread writing results in file order.
• A work queue built on a mutex and a condition_variable, a std::map that puts results back in order, a counting_semaphore that bounds the blocks in flight, and RAII file handles.
• Every version keeps the same 683,106 rows and removes 6,992,695 (91%), with byte-identical output (same SHA-256).
• A written review lists 11 mistakes in the original and one open design question, the 91% row-drop rule. A byte-for-byte test on a 12-line sample and a benchmark script check the output and record the runs.

Result

The single-threaded original took 447.5 s as a Debug build. The rewrite takes 2.2 s, 203× faster, in 18–32 MB of memory, and keeps the same 683,106 rows.

Lead Architect for a five-person team at ASA DataFest 2026. I set up the repository, wrote Python scripts and the first C++ cleaner for the 1.47 GB file.

Problem

The dataset was a 1.47 GB file, too big for Excel or Access, and a team of five had to analyze it together.

Approach

Get the team’s tools in place first: a shared repository, development environments, roles and a practice schedule. Then clean the full file in C++ and analyze it with short Python scripts.

Implementation

• Led the technical design of the team’s data analysis.
• Set up the GitHub repository, development environments, roles and a practice schedule.
• Wrote the original single-threaded C++ cleaner for the 1.47 GB file (see Console File Cleaner).
• Four small Python scripts with pandas, seaborn and matplotlib: a Top-20 diagnosis-groups chart and summary, and value counts and top-category charts across 8 patient columns.

Result

A cleaned file small enough to analyze, a Top-20 diagnosis-groups chart and value counts across 8 patient columns.

A personal JavaFX desktop app for tracking an 8-week schedule, built with Java 21, SQLite and JUnit.

Problem

Plan an 8-week schedule session by session and keep weekly notes in a Markdown file.

Approach

A small desktop app for my own use: JavaFX for the interface, SQLite for storage, and notes appended to a plain Markdown file.

Implementation

• A session grid and a week editor, written in JavaFX with no FXML.
• Notes are appended to a weekly-notes.md file.
• 18 JUnit tests across about 1,580 lines of Java.

Result

A personal desktop app with 18 JUnit tests, about 1,580 lines of Java.

An alpha .NET 10 tool that records what AI coding agents do: the problems they find and the fixes they make.

Problem

I wanted to see what AI coding agents actually do in a session, not just the result they hand back.

Approach

Log the problems and fixes to SQLite as agents work. Expose the log to agents as an MCP server, so they record it themselves, and show it in a desktop dashboard.

Implementation

• A core library that stores records in SQLite through Dapper.
• A command-line tool that is also an MCP server, exposing log_issue, log_exception, log_fix and list_open_issues.
• A WPF desktop dashboard with charts.
• Capture of Claude Code hook events.
• 165 xUnit tests across about 9,600 lines of C# plus XAML.

Result

An alpha tool (phase 5 of 6) with 165 xUnit tests, built with Claude Code to see what AI agents actually do.

A menu-driven birthday database in C++, built on a template binary search tree.

Problem

Keep a collection of birthdays that can be searched, edited and saved to a file.

Approach

Store the entries in a binary search tree written as a template, and save them to a file between runs.

Implementation

• A node class, a tree class and a binary search tree built in layers, with nodes held by std::shared_ptr.
• Custom exceptions for entries not found and violated preconditions. The Person class overloads the comparison operators.
• A menu to add, modify, remove and search entries, list everyone born in a given month, and save to or load from a file.

Result

Practice combining a template data structure, file I/O and a menu-driven program.

Six sorting algorithms in C++, plus a timing program that runs them on random integers.

Problem

Compare how different sorting algorithms behave as the input grows.

Approach

Write each sort and run them on random integers at four sizes, from 10k to 1M, skipping the O(n²) sorts at 1M.

Implementation

• Selection, bubble, insertion, merge, quick (median-of-three pivot) and radix sort.
• A SortLab class that holds the data in a dynamic array and follows the rule of five: copy and move constructors and assignment, and a destructor.
• A timing program that uses clock() on 10k, 50k, 100k and 1M random integers.

Result

Practice with how each sort works, and with managing dynamic memory in C++.

Linked lists, stacks, queues and a bag as C++ templates, plus a palindrome checker and a rainfall-statistics program.

Problem

Build the basic linked data structures as reusable templates and use them in small programs.

Approach

Write each structure as a C++ template so the same code works for any type, then use the structures in small programs.

Implementation

• A linked list, array list, linked stack, array stack, a queue built on the linked list, and a linked bag, all as templates.
• A palindrome checker that compares a queue of characters against a stack of characters.
• A rainfall-statistics program built on its own linked list.
• Recursion exercises, including recursive binary search.

Result

Practice with templates, pointer-based structures and recursion.

A C++ program that checks whether two binary search trees hold the same values, using recursive in-order traversals.

Problem

Check whether two binary search trees hold the same values, whatever their shape.

Approach

Traverse each tree in order, recursively, into a vector, and compare the two vectors.

Implementation

• A recursive in-order traversal of each tree into a vector.
• A comparison of the two vectors, so the shape of the tree doesn’t matter.

Result

Practice with recursion on binary search trees.

Exercises from an intro Java course: a Luhn credit-card number check, a loan calculator, and JavaFX tax and investment calculators.

Problem

Practice the basics of Java with small programs.

Approach

Small programs for the basics, with a JavaFX interface for the tax and investment calculators.

Implementation

• A credit-card number check (Luhn).
• A console loan calculator.
• JavaFX tax and investment calculators.

Result

Core Java syntax and a small JavaFX interface.