# Sara Kay — Project Inventory

> Every project on sarakay.me, in one machine-readable file. Sara Kay is an
> applied AI engineer (production LLM systems, RAG, agentic workflows, AI data
> governance) in Boston, MA. This file is generated from the same data as
> https://sarakay.me/projects.html and https://sarakay.me/case-studies.html.

Related machine-readable resources:
- Site guide for AI agents: https://sarakay.me/llms.txt
- Career history (CSV): https://sarakay.me/sara_kay.csv
- Talks and conference history: https://sarakay.me/talks.html
- Case study deep dives have Markdown twins at case-studies/<slug>.md

## Agentic Tools

*AI tools that ship with a Claude Code skill, each with a hard line the model is not allowed to cross: the inbox stays read-only, a human writes the Wikipedia article, nothing gets deleted.*

### Internet Historian
*status: active · metric: 24/7 politely retrying*

A patient Wayback Machine archiver. It treats a busy Internet Archive as weather rather than failure and retries until your pages are actually preserved.
- **Tech**: Python, SQLite, launchd, Wayback Machine, PyPI, Claude Code skill
- **view on GitHub**: https://github.com/ssskay/internet-historian
- **read more**: https://sarakay.me/case-studies/internet-historian.html

### Wikipedia Gap Finder
*status: active · metric: 212 women documented*

Finds people from underrepresented groups missing from English Wikipedia and fact-checks every claim into a research dossier. The AI does the detective work; a human writes the article.
- **Tech**: Python, MediaWiki API, Wikidata, Claude Code skill, No API keys required
- **view on GitHub**: https://github.com/ssskay/wiki-gap-finder
- **try the demo**: https://sarakay.me/demos/skills/#gapfinder

### Inbox Catalog
*status: active · metric: read-only by design*

A local-first, read-only engine that turns order emails into a structured catalog of what you own, what's coming, and what you spent.
- **Tech**: Python, IMAP, Claude Code plugin, Local-first, PyPI
- **view on GitHub**: https://github.com/ssskay/inbox-catalog

### MyFigureCollection API
*status: active · metric: 8 years of waiting for API v4*

Unofficial API for MyFigureCollection.net, written after eight years of an officially promised API v4 that never shipped. A Python library plus a local MCP server with 12 tools, covering item search, JAN barcode lookup, partner shop prices, collections, and clubs.
- **Tech**: Python, MCP server, curl_cffi, Web scraping, PyPI
- **view on GitHub**: https://github.com/ssskay/myfigurecollection-api

### Hamster Wheels
*status: active · metric: 17 agents craft + claude mcp*

17-agent AI system for personal life management, built in Craft on Claude MCP. Each agent owns one life domain: triage, career, finance, events.
- **Tech**: Craft, Claude, MCP, Agentic workflows
- **read the case study**: https://sarakay.me/case-studies/hamster-wheels.html

### Hamster Sweep
*status: shipped · metric: auto intel sweeps*

Background intelligence sweeps for two domains: Japanese reading practice from live news, and structured job-opportunity cards. Results land straight in Craft via MCP.
- **Tech**: Python, Anthropic SDK, web_search tool, Craft MCP

## Desktop Pet Rescue

*Cute Windows-only desktop apps, given a second life on native macOS. The original developers did the hard, lovely work; I do the porting, credit them prominently, and send fixes back upstream.*

### Yaha-Pet for macOS
*status: active · metric: 3 pets on your desktop*

Native macOS port of gitChara-dot's Yaha-Pet: Usagi, Hachiware, and Chiikawa roam the desktop. Two bug fixes offered back upstream in an open PR.
- **Tech**: Python, PyQt6, PyInstaller, macOS
- **view on GitHub**: https://github.com/ssskay/Yaha-Pet
- **read the rescue story**: https://sarakay.me/case-studies/desktop-pet-rescue.html

### Desktop Chiikawa for macOS
*status: active · metric: 5 skins · 5 languages*

Native macOS port of CookieElmo's wellness-reminder pet from itch.io, rebuilt on the official Godot runtime. It reminds you to drink water.
- **Tech**: Godot 4.4, GDScript, Python, macOS
- **view on GitHub**: https://github.com/ssskay/desktop-chiikawa-macos
- **read the rescue story**: https://sarakay.me/case-studies/desktop-pet-rescue.html

### Desktop Pet Rescue
*status: active · metric: 2 pets rescued so far*

A porting practice: adopting cute Windows-only desktop apps and giving them a second life on native macOS. Two Chiikawa pets so far, with prominent credit to the original developers and fixes offered back upstream, including an open PR for Mac support.
- **Tech**: Python, PyQt6, Godot 4.4, GDScript, macOS
- **read the case study**: https://sarakay.me/case-studies/desktop-pet-rescue.html

## Data & Pipelines

*Pipelines, sync engines, and the observability habits that follow me home from work.*

### Anime Analyst
*status: active · metric: 25k+ anime entries*

Natural-language query system over 25k+ anime entries: PostgreSQL, a local ReAct SQL agent, and a Streamlit frontend. CMU capstone, presented at Anime Boston 2026.
- **Tech**: Python, PostgreSQL, LangChain, Streamlit, Ollama, ReAct agents
- **read the case study**: https://sarakay.me/talks/living-in-the-database.html

### Twitter Archive Pipeline
*status: shipped · metric: NLP self-analysis*

ETL pipeline that turns Twitter/X export files into a searchable dataset with NLP sentiment and topic classification, built for longitudinal self-analysis.
- **Tech**: Python, Pandas, spaCy, DuckDB, Dagster
- **the backstory**: https://www.linkedin.com/feed/update/urn:li:share:7343248219474149377

### Digital Wellness Through Data
*status: shipped · metric: 26,720 likes, finally read*

Interactive case study mining a decade of personal social media data (26,720 Twitter likes, 4,500 saved YouTube videos, 500+ LinkedIn connections) to reveal what the algorithm sees. CMU data mining final project, published as a scrollytelling site.
- **Tech**: Python, Pandas, HuggingFace Transformers, Scikit-learn, Statsmodels, React, Recharts
- **explore the study**: https://digital-wellness-through-data.vercel.app
- **read more**: https://sarakay.me/case-studies/digital-wellness-through-data.html

### Earditor
*status: active · metric: blank beats wrong*

Evidence-based music tagger for macOS that cross-examines three independent witnesses (Shazam, AcoustID + MusicBrainz, iTunes) into six scored signals and a verdict tier per file. Treats a cover as a first-class verdict instead of a wrong catalog match, and never writes a tag it can't prove.
- **Tech**: Python, Flask, Shazam, AcoustID + MusicBrainz, iTunes Search API, SQLite
- **view on GitHub**: https://github.com/ssskay/earditor
- **try the demo**: https://sarakay.me/demos/earditor

### MangaRelease
*status: active · metric: 15+ publishers, daily*

Automated open-data release calendar for licensed English manga, manhwa, manhua, and webtoons, forked from LNRelease's engine with a rebuilt VIZ scraper and historical calendar backfill. Daily and weekly GitHub Actions scrape 15+ publishers, auto-commit, and auto-deploy the site with no human in the loop.
- **Tech**: Python, GitHub Actions, Web scraping, robots.txt compliance, Static site, Open data
- **view on GitHub**: https://github.com/mangarelease/mangarelease.github.io
- **see the calendar**: https://mangarelease.github.io

### Media Collection Sync Engine
*status: in-progress · metric: 1,600+ records synced*

Media-collection sync engine keeping 1,600+ records aligned between a REST API and cloud storage, using GPT-4o fuzzy matching with a human-in-the-loop review UI.
- **Tech**: Python, Flask, OpenAI (gpt-4o + embeddings), MEGAcmd, REST API

### Repo Observatory
*status: active · metric: 14 days is all github remembers*

GitHub keeps only 14 days of repository traffic history, so a private pipeline snapshots every repo daily into permanent CSVs with a self-updating README dashboard. SRE habits, applied to side projects.
- **Tech**: GitHub Actions, Python, matplotlib, CSV

## Fandom & Community

*Tools that exist because a community needed them, or because I did.*

### Which Odyssey?
*status: active · metric: 2,800 years of retellings*

Recommendation quiz that matches readers, viewers, and listeners to the right Odyssey adaptation. 24 static pages covering 17 versions (translations, film, novels, albums, games) with structured data and FAQ content built for search and AI answers.
- **Tech**: Static site, Cloudflare Workers, SEO, JSON-LD, GA4
- **take the quiz**: https://whichodyssey.com/

### Which Adaptation?
*status: active · metric: one engine many stories*

Hub for a family of adaptation guides built on one recommender engine and one catalog model. Every story gets its own property, quiz, and comparison matrix, with structured data and FAQ content built for search and AI answers.
- **Tech**: Static site, Cloudflare Pages, SEO, JSON-LD, Recommender
- **visit the hub**: https://whichadaptation.com/
- **which odyssey?**: https://whichodyssey.com/

### Uma Musume Q&A Bot
*status: shipped · metric: RAG discord bot*

Citation-first RAG Discord bot for Uma Musume game mechanics, answering with page-level source attribution from a 200+ page community reference doc.
- **Tech**: Python, LangChain, OpenAI Embeddings, FAISS, Discord.py
- **read the case study**: https://sarakay.me/case-studies/uma-musume-bot.html

### K-Pop Button Maker
*status: shipped · metric: face AI con buttons*

Desktop GUI that turns K-pop idol photos into convention-ready 1.5" button designs: automatic face detection, production-spec cropping, bleed margins.
- **Tech**: Python, CustomTkinter, InsightFace, OpenCV, Pillow

### Campus Freebies
*status: active · metric: $0 price of everything listed*

An agent-maintained catalog of free student and alumni perks: searchable list, campus map, weekly link checker. Live for Dartmouth and CMU; forkable for any school.
- **Tech**: Astro, Supabase, Vercel, GitHub Actions
- **fork it for your school**: https://github.com/ssskay/freestuff-template

## Odds & Ends

*Research, writing, and one very effective decluttering app.*

### Data Gremlin Go Brr
*status: active · metric: live newsletter*

Substack newsletter on AI tools and data systems for practitioners: the intersection of technical work with everyday life.
- **Tech**: Substack, Technical writing, Data storytelling
- **read the newsletter**: https://newsletter.datagremlingobrr.com

### GPT-3 Mobile Gaming Research
*status: legacy · metric: pre-ChatGPT Dartmouth thesis*

Dartmouth senior thesis running experiments on GPT-3 months before ChatGPT's release, analyzing mobile-gaming monetization and player behavior.
- **Tech**: GPT-3, Python, NLP, Sentiment Analysis, Machine Learning, Quantitative Social Science

### Home Zone Tracker
*status: shipped · metric: 15 zones tracked*

React SPA for tracking apartment decluttering zone-by-zone: 15 zones, sprint mode, deadline countdown. Built for real use before a trip.
- **Tech**: React 19, TypeScript, Vite, Tailwind, lucide-react
- **try the demo**: https://sarakay.me/demos/zones

### Osier
*status: active · metric: Fn+Ctrl tap, speak, tap*

Notch-native dictation for macOS, and a hard fork of OpenSuperWhisper. Tap Fn and Control, speak, tap again, and the transcription lands in whatever field was already focused. The upstream project owns the engine; this fork reshapes the trigger and the interface.
- **Tech**: Swift, SwiftUI, macOS, Whisper, On-device ML
- **view on GitHub**: https://github.com/ssskay/osier

## Professional & Extended Case Studies

*Work systems and practices documented as case studies (not listed on the projects page).*

### Keysensus
*status: active · metric: in prod shipped · IBM*

Context-aware data-governance service at IBM (HashiCorp). Applies policy decisions using business and service context — instead of static field-only matching — across distributed production pipelines, with an auditable draft → review → publish → revoke lifecycle. The control layer that determines what data is safe for downstream AI and analytics systems to use.
- **Tech**: Data governance, Distributed systems, Production deployment, IBM, HashiCorp

### ETL Pipeline Optimization
*status: shipped · metric: 54% faster ETL · Moody's*

Modernized high-throughput data pipeline at Moody's Analytics using PySpark and Databricks, reducing processing time by 54% while handling larger datasets — the data-velocity foundation downstream analytics and AI workloads depend on.
- **Tech**: PySpark, Databricks, ETL, Performance optimization, Moody's Analytics

### Sanctions Compliance System
*status: shipped · metric: $1.2M revenue · Moody's*

Built AWS Bedrock-powered sanctions compliance system at Moody's Analytics that generated $1.2M in revenue. Secure, compliant production AI for financial regulatory workflows — high-security AI tooling for finance that customers paid for.
- **Tech**: AWS Bedrock, LLMs, Compliance, Revenue generation, Moody's Analytics

### LangChain Citation System
*status: shipped · metric: 92% citation accuracy*

Developed LangChain-powered RAG citation system at Moody's Analytics using GPT-4 Turbo, achieving 92% citation accuracy. Solved the hallucination problem with page-level source verification — LLM evaluation and grounding work, before it had a job title.
- **Tech**: LangChain, GPT-4 Turbo, RAG pipelines, Citation accuracy, Moody's Analytics

### Gremlin Goods
*status: shipped · metric: OPEN for commissions*

Maker studio producing laser-cut items, sublimation prints, acrylic charms, and button sets, with automated design-to-production workflows. Commissions open for conventions, events, nonprofits, and artist collabs. Charity donations (GDQ, Frost Fatales) stay at the core.
- **Tech**: Glowforge, Silhouette Cameo, Sublimation printing, Button maker
- **view the gallery**: https://sarakay.me/maker.html

*Generated by tools/gen_md.py from data/projects.csv. Canonical HTML: https://sarakay.me/projects.html*