Projects

Home energy upgrades are worth thousands in rebates that almost nobody claims, because working out which ones you qualify for means reading your own appliance labels and utility bills. EcoStep reads them for you.

A model was retired mid-build

The production Gemini model the vision pipelines ran on was retired while the project was still being built. I replaced it with a 4-model automatic fallback chain and a 9-category error taxonomy across both pipelines, appliance-label OCR and utility-bill parsing, which took single-model outages off the table entirely.

Five engines, six routes

Rebate matching, affordability simulation, outage-resilience planning, group-buying, and proactive coaching, across 6 REST API routes. Benchmarked at 9,600+ evaluations per second, averaging 0.104ms per call, with 8 of 8 unit tests passing.

What was broken

Three classes of app-wide UI bug: a dark-mode toggle that was silently ignored, 17 invalid Tailwind color tokens, and 17 controls that failed contrast. I fixed all three, added a motion layer across 6 components, and shipped with zero TypeScript errors across 13 routes.

• Next.js
• React
• TypeScript
• Tailwind CSS
• Supabase
• Gemini API
• Vercel

Built with Abeni Rodriguez, Emma Stefanini

Source on GitHubWriteup on Devpost

Writing a professional email in a second language is slow, and most tools either translate too literally or flatten the English you already wrote correctly. Bolo takes text, speech, or a screenshot of an email and returns a drafted English reply.

The measured result

Drafting time dropped from 30 minutes to 4 across 50+ generated emails, an 87% reduction.

Code-switching without losing intent

People writing in a second language mix English words, phrases, and whole sentences into their native text. Prompt flows that passed 35 generation tests preserve those rather than round-tripping them back out through translation.

What an adversarial QA pass found

21 verified bugs, fixed, with 126 tests passing afterwards. The fixes were prompt-injection detection, Unicode-aware OCR cleanup, runtime API validation, stale async guards, and an empty-response fallback.

• Next.js
• React
• TypeScript
• Tailwind CSS
• Tesseract.js
• Gemini API
• OpenAI API
• Vercel

Try BoloSource on GitHub

Recipe sites want you to already have the ingredients. Cooked! goes the other way: type in what is actually in your fridge and it ranks recipes by how much of each one you can make right now.

Ranking, not just matching

A tiered ingredient-matching algorithm with coverage-based scoring ranks 12+ recipes per query, rather than dumping back everything that happens to contain one ingredient you own.

Cutting redundant calls

In-memory caching and API error handling reduced redundant external API calls by 73.6% across overlapping query benchmarks.

People type messily

Nobody enters ingredients consistently. '2 cups', '2c', and 'two cups' all show up. Real-time validation, unit normalization, and unit suggestions resolve that before anything reaches the matcher.

• Python
• Flask
• Next.js
• TypeScript
• REST APIs
• TheMealDB API
• Gemini API

Built with Natalie Reese

Source on GitHubWriteup on Devpost

A birthday present for my best friend. A mobile-first soundboard styled after a Stream Deck: sixteen tactile keys in a 4×4 grid, one sound each, no build step and no dependencies.

No framework, deliberately

Vanilla HTML, CSS, and JavaScript. Nothing to install, nothing to build. It registers a service worker and installs as a PWA, so it keeps working with no connection at all, which matters for something meant to be opened on a phone at a party.

iOS is the hard part

Safari suspends the Web Audio context until a real user gesture unlocks it, then suspends it again when the page is backgrounded. app.js carries a dedicated recovery path for that, and the repo ships a separate diagnostics page for testing on the phone itself rather than in a desktop simulator.

Swappable without touching code

The sixteen keys are built from a manifest.json that maps files to slots, with an optional colour shade per key. Changing the whole sound set means editing one JSON file and reloading.

• JavaScript
• HTML
• CSS
• Web Audio API
• PWA

Source on GitHub

Given a network and a list of connections that fail one after another, how connected does the network stay at each step? This answers that in O(log n) average time per operation.

Running the problem backwards

Union-Find only merges. It has no way to split a set, so sequential edge removal is the one thing it cannot do directly. Reverse-deletion sidesteps that: replay the removals backwards as insertions, rebuilding the graph bottom-up and tracking size-weighted component merges at each step to recover the connectivity score at every stage.

Where the log n comes from

Union-Find with both path compression and union by rank. Either alone is worse; together they give the O(log n) average.

• Java
• Algorithms

Source in my coursework repo

A team build of a RISC-V CPU system-on-chip that runs video games on an FPGA, run through IEEE at UCF. Hardware is the half of computer engineering a software portfolio usually leaves out.

The design

A 5-stage pipeline with VM/0 fetch, built as a team through the university's IEEE chapter.

My part

Coding, debugging, synthesis, and implementing hardware modules: the ordinary work of getting RTL to behave on real hardware rather than only in simulation.

• Verilog
• FPGA
• RISC-V
• RTL Design

IEEE at UCF

A card game simulator in C that reads a deck from a plain text file, deals it out, and tracks every player's hand. Written while learning what linked lists are actually for.

Where the work was

Sorted insertion, pile reversal, and transfers between the draw, player, and discard piles, all on hand-rolled linked lists, with dynamic memory allocation and file I/O underneath.

• C