Portfolio & Projects

Real-world builds across Generative AI, Machine Learning, and shipped products

Generative AI

LLM applications, RAG systems, and AI-powered tools built with Claude, OpenAI, and Deepgram

3 projects
AI Application

Pass the Mic: AI Recording Studio

Pass the Mic is a voice-to-text app built for creators who think out loud. Whether you're freestyling bars, recording a podcast, or dictating a book chapter, hit record and get a clean transcript instantly. Switch between Lyrics, Podcast, and Book modes for a tailored experience. Podcast mode includes speaker diarization, guest name mapping, episode metadata, and transcript export. Book mode supports chapter metadata for publish-ready output. All sessions are saved to Google Drive - no database needed.

The Challenge

Writers and performers often freestyle or workshop lyrics out loud but lose their best lines. Podcasters need speaker-labeled transcripts for show notes and SEO. Authors need a fast way to convert spoken words into publish-ready text.

The Solution

Built a multi-mode recording studio powered by Deepgram's Nova-3 model with smart formatting, speaker diarization, Google Drive storage, and video support with automatic audio extraction.

View on GitHub

Key Features & Results

  • Deepgram Nova-3 speech-to-text integration
  • 3 modes: Lyrics, Podcast, and Book
  • Speaker diarization with guest name mapping
  • Google Drive storage with OAuth sign-in
  • Video support - MP4, MOV, AVI, MKV
  • Audio trimming and transcript export
Web Application

AI Website With Integrated Chatbot

A complete website with an AI-powered chatbot widget - just like the ones on real company websites. When a visitor clicks the chat icon, they get instant AI-powered responses. The chatbot uses RAG (Retrieval-Augmented Generation) to read from your own text documents so it can answer questions about YOUR specific content. Includes a marketing site with navigation, hero section, features, and footer, plus a floating chat widget with conversation memory so the AI remembers context throughout the chat.

The Challenge

Businesses need intelligent chatbots that understand their specific products and services, not generic AI responses.

The Solution

Built a full-stack web application with an embedded AI chatbot powered by OpenAI, using RAG to ground responses in the business's own content, with conversation memory for natural multi-turn interactions.

View on GitHub

Key Features & Results

  • Full marketing website with nav, hero, features, footer
  • Floating chat widget with open/close toggle
  • OpenAI-powered intelligent responses
  • RAG - chatbot reads from your own documents
  • Conversation memory across the chat session
AI Application

AI Job Search System

No recruiter. No network. No problem. Five peer skills, each mapped to one job-to-be-done: discover open roles from Greenhouse/Lever/Ashby APIs, apply with stage-adaptive tailored resume and cover letter PDFs, generate interview prep briefs with STAR story mappings, review your pipeline with stale flags and pacing checks, and a deliberate submit stub that keeps submission human-only. The system reads from a single knowledge/ directory (your profile, master resume, story bank, target companies) so anyone can fork it and make it theirs. Role-agnostic - works for AI Engineer, Product Manager, UX Designer, Cybersecurity Analyst, or anything else.

The Challenge

Job searching is a repetitive, high-effort process scattered across dozens of tabs. Every application needs a tailored resume, custom cover letter, and interview prep - and most people either use a generic resume or burn out customizing each one.

The Solution

Built a 5-skill Claude Code system with direct ATS polling, stage-adaptive resume positioning (early/growth/enterprise variants per bullet), STAR story bank discipline (flags gaps instead of inventing stories), Chrome headless PDF rendering with zero dependencies, and a forkable knowledge/ directory that separates user data from skill logic.

View on GitHub

Key Features & Results

  • 5 specialized Claude Code skills
  • Direct Greenhouse/Lever/Ashby API polling
  • Stage-adaptive resume tailoring
  • Grade-A PDF resume + cover letter generation
  • STAR story bank with honest gap flagging
  • Pipeline tracker with stale detection
  • Role-agnostic - any job title works
  • Zero Python dependencies - stdlib only

Machine Learning

Classical ML across recommendation systems, classification, clustering, and predictive modeling

4 projects
Machine Learning

Amazon Product Recommendation System

Built and compared three recommendation approaches on a 65,000-record subset of Amazon's electronics ratings dataset (sourced from a 7.8M-record Kaggle dataset). The rank-based model handles cold-start users with no history. KNN collaborative filtering (user-user and item-item) personalizes recommendations based on similar users and products. SVD matrix factorization performs best on sparse interaction data by learning latent factors. All models evaluated with RMSE, Precision@k, Recall@k, and F1@k using scikit-surprise.

The Challenge

Amazon's catalog contains millions of products. Cold-start users have no history, while active users need personalization that scales across sparse rating data. No single algorithm handles both cases well.

The Solution

Built a three-model pipeline: rank-based baseline for cold-start, KNN collaborative filtering for personalized recommendations, and SVD matrix factorization for sparse-data performance. Used scikit-surprise for model training and evaluation across RMSE, Precision@k, Recall@k, and F1@k.

View on GitHub

Key Features & Results

  • Rank-based popularity model for cold-start users
  • KNN collaborative filtering (user-user and item-item)
  • SVD matrix factorization, best performance on sparse data
  • Evaluated with RMSE, Precision@k, Recall@k, F1@k
  • Built on 65K-record subset of 7.8M Amazon electronics ratings
  • Powered by scikit-surprise
Machine Learning

Customer Personality Segmentation

Applied K-Means clustering on 2,240 customer records spanning 30 attributes: demographics, spending across 6 product categories, and purchase channel preferences (web, catalog, in-store). Used the elbow method to determine optimal k=4 and silhouette analysis to validate cluster quality. Each of the 4 segments has a distinct behavioral profile: High-Value Enthusiasts (premium spenders, catalog/store preference), Deal-Seeking Families (household focus, discount-driven), Opportunity Segment (web-active but low conversion), and Balanced Shoppers (mid-income, omnichannel). Key insight: income strongly drives spending, but web traffic alone doesn't predict conversion.

The Challenge

Marketing teams broadcast the same message to every customer, leaving segment-specific opportunities on the table. Without behavioral groupings, personalization is guesswork.

The Solution

K-Means clustering with elbow method for optimal K and silhouette analysis for cluster validation. Profiled each cluster across income, spend habits, family structure, and channel preference to generate actionable segment descriptions.

View on GitHub

Key Features & Results

  • K-Means clustering with elbow method (optimal k=4)
  • Silhouette analysis for cluster quality validation
  • 4 named segments with distinct behavioral profiles
  • High-Value Enthusiasts: premium spenders, catalog/store preference
  • Deal-Seeking Families: household focus, discount-driven
  • Opportunity Segment: web-active, low conversion rate
  • Balanced Shoppers: mid-income, omnichannel buyers
Machine Learning

Potential Customers Prediction

Binary classification pipeline on 4,612 leads from ExtraaLearn (an ed-tech platform) with a 29.9% conversion rate. Compared Decision Tree baseline against a tuned Random Forest with class weighting to handle the imbalance. Optimized for recall to minimize missed conversions rather than false positives. Final model: 83% accuracy, 85% recall, F1 of 0.76. Feature importance analysis revealed that time spent on the website (0.301), first interaction source (0.281), and profile completion (0.200) are the top three conversion drivers. Referral leads convert at 67.7%, website first-touch at 45.6%.

The Challenge

Ed-tech sales teams waste budget chasing low-probability leads. With only 30% of leads converting and class imbalance in the data, a naive model would simply predict non-conversion every time.

The Solution

Decision Tree baseline compared against a tuned Random Forest with class weighting ({0: 0.3, 1: 0.7}) and GridSearchCV hyperparameter optimization. Recall-optimized to prioritize finding real converters over minimizing false positives.

View on GitHub

Key Features & Results

  • Decision Tree baseline + tuned Random Forest with GridSearchCV
  • Class weighting to handle 29.9% conversion rate imbalance
  • 83% accuracy, 85% recall, F1 score of 0.76
  • Top signal: time on website (importance: 0.301)
  • First interaction source: 0.281 importance
  • Referral leads convert at 67.7% vs. website at 45.6%
Machine LearningHackathon

Travel Experience Prediction

Hackathon project predicting passenger satisfaction (satisfied / not satisfied) on Japan's Shinkansen bullet train network using 94,379 training records and 25 features including demographics, trip details, and 14 ordinal service ratings. Progressed from Decision Tree and Random Forest baselines to an ensemble of gradient boosting models (XGBoost, LightGBM, CatBoost) combined via a soft-vote VotingClassifier. CatBoost final model hit 95.69% leaderboard accuracy, good for 2nd place. Feature importance analysis found onboard entertainment (50.2%) and seat comfort (20%) as the top satisfaction drivers. Punctuality had minimal impact on satisfaction.

The Challenge

Predicting subjective satisfaction from service rating data requires a model that captures non-linear relationships across 14 ordinal features. Simpler models plateau quickly on this type of data.

The Solution

Gradient boosting ensemble: XGBoost, LightGBM, and CatBoost individually tuned with GridSearchCV, then combined into a soft-vote VotingClassifier. CatBoost (depth=8, 10,000 iterations, lr=0.01) achieved the best standalone performance at 95.69% accuracy.

View on GitHub

Key Features & Results

  • 2nd place hackathon finish, 95.69% leaderboard accuracy
  • Gradient boosting ensemble: XGBoost + LightGBM + CatBoost
  • Soft-vote VotingClassifier combining all three models
  • GridSearchCV hyperparameter tuning per model
  • 94,379 training records, 25 features
  • Top driver: Onboard Entertainment (50.2% feature importance)
  • Seat Comfort second at 20%, punctuality had minimal impact

Products

Shipped products live in the real world with real users

1 project
Product Launch

Amiellevia: Global Lesbian Social App

Amiellevia is the all-in-one platform built by and for lesbian and queer women. Instead of juggling separate apps for dating, making friends, finding events, and connecting with community, everything lives in one place. AI-powered compatibility matching goes beyond photos to find meaningful connections. Users can join or create communities by city, interest, or identity, discover and host events with built-in ticketing and QR check-in, and message anyone directly. Live on the App Store and Google Play since December 2025.

The Challenge

Lesbian and queer women are forced to use mainstream apps that weren't designed for them, scattered across multiple platforms for dating, friendship, events, and community. Moving to a new city or traveling means starting from scratch with no way to find your people.

The Solution

Built a five-pillar mobile platform -Dating Pool, Friend Zone, Communities, Events, and Direct Messaging -with AI-powered matching, paid event ticketing, community management tools, and virtual date scheduling, all in a safe, moderated space.

View Project

Key Features & Results

  • AI-powered compatibility matching beyond photos
  • 5 pillars: Dating, Friends, Communities, Events, Messaging
  • Join or create communities by city, interest, or identity
  • Host events with paid ticketing and QR code check-in
  • Virtual date scheduling with video call integration
  • Live on iOS App Store and Google Play

More Projects

Additional work, experiments, and builds

Your First AI Claude Workflow

A skill system that turns repeatable processes into one-line slash commands using Claude Code Skills and markdown files.

Claude CodeAI WorkflowsMarkdown
View on GitHub

AI Personal Portfolio Builder

A ready-to-deploy portfolio template with an OpenAI RAG chat widget and 4 Claude Code skills for generating bios, copy, and social profiles.

OpenAIRAGClaude Code
View on GitHub

Want Me to Build Something For You?

Websites, apps, logos, branding - let's bring your vision to life