Arman
Sarrafi

1. 01

   Direct, lightweight data collection

   Inventory used to come from browser-based scrapers. I rebuilt the scanner to call dealers' data feeds directly over HTTP, so it is simpler to run and much lighter to deploy: the container shrank from 11.8 GB to 1.09 GB, and it now covers 454 of 459 active dealers.

2. 02

   Search in plain language

   Shoppers describe what they want and get matches from semantic search (pgvector) combined with normal filters like price and distance. If semantic search is ever unavailable, it falls back to standard filtering automatically.

3. 03

   Web, dealer portal, and iOS on one API

   A Flask REST API powers the website, a dealer portal, and a native SwiftUI iPhone app with ZIP-radius search, saved cars, side-by-side comparison, and a dealer map. Cached listing requests return in about 5 ms.

4. 04

   An assistant that sticks to the facts

   The built-in car assistant runs on Claude or a local Ollama model and treats listing and web text as untrusted input. Stripe handles billing for premium features.

5. 05

   Built to stay reliable

   Over 5,300 automated tests, GitHub Actions CI, guarded releases and deploys, and nightly data-quality checks keep about 215K listings accurate.

Sample listing photos by Renée Kools and Autowizja (CC BY 4.0), Trop86, RichardSummersault and Retired electrician (CC0), and Bull-Doser (public domain), via Wikimedia Commons. Cropped, plates blurred. Full credits

1. 01

   Finds the right jobs

   Zumai scans employer boards every day through the public Greenhouse, Lever and Ashby APIs. Local embeddings shortlist the top 100 matches, then Claude scores and ranks them against your profile.

2. 02

   Applies carefully

   It drafts tailored resumes and cover letters and can apply through browser automation (Playwright). It runs as a dry run by default, never applies to the same job twice, and never guesses an answer to a legal question.

3. 03

   Respects the sites it reads

   It uses official job-board APIs instead of scraping pages, follows robots.txt with no exceptions, and rate-limits itself per site.

4. 04

   AI costs stay predictable

   Every AI call is logged with its model, token count and cost, and a daily spending cap keeps usage in check.

5. 05

   Works with other AI tools

   An MCP server lets outside AI agents use Zumai's tools, with a separate access key for each user.

1. 01

   Answers from your own courses

   Each student connects their own Canvas account. SchoolCenter syncs and indexes it, then answers questions with citations back to the original material.

2. 02

   Your week in one place

   Assignments from every course sync from Canvas into a single list, grouped by what is overdue, due soon and later, so nothing gets missed.

3. 03

   Study tools built from your materials

   Flashcards, summaries, quizzes and an exam mode are generated from your own course documents, with the sources shown next to each answer.

4. 04

   Private by default

   AI runs locally first, through Ollama or Apple Intelligence. Cloud AI is used only after the student reads a notice and chooses to continue.

5. 05

   Built for academic integrity

   The browser extension stays off quizzes, submissions and gradebooks, and there is no endpoint that writes answers for a student.

6. 06

   Hardened and well tested

   A seven-phase security review followed by fixes, backed by more than 8,300 automated tests, keeps each student's data separate.

7. 07

   Also on iPhone and Mac

   An early SwiftUI companion app brings study tools like spaced repetition to Apple devices.

1. 01

   The question

   Where in Orange County, California do drivers most need new EV chargers? The goal is a repeatable, data-backed ranking instead of guesswork.

2. 02

   The data

   State EV registrations, federal charging-station data, traffic counts, Census ZIP centers, and OpenStreetMap amenities such as restaurants and malls, cleaned and combined in Python with pandas.

3. 03

   The ranking

   Each ZIP code is scored by how many EVs it has per charging station, then weighted by nearby traffic and amenities.

4. 04

   The map

   An interactive folium map shows the biggest gaps alongside existing chargers and nearby places, so each gap can be checked at a glance.

1. 01

   A Pi that talks to the truck

   A Raspberry Pi 5 running LineageOS (Android 16) acts as the phone for my RAM truck's Uconnect 5 head unit. A Linux USB gadget set up through configfs completes the Android Open Accessory handshake over a wired USB connection.

2. 02

   The whole tablet on the truck screen

   A mirror app (AAStream) projects the full Android interface onto the truck display instead of the stock Android Auto launcher. A setup script applies the root and permission grants (Magisk) so it works hands-free.

3. 03

   Reliable on every boot

   Boot hooks re-arm the USB accessory mode each time the Pi starts, a watchdog keeps wifi up, and small collectors record evidence when something goes wrong. Each script does one job, can be re-run safely, and logs to its own file.

4. 04

   Documented like a real project

   Runbooks, a conventions file, and a list of gotchas keep it reproducible. Bench testing turned up bugs that I fixed, and logs showed that one outage was a power cut, not software. 55 commits so far.

5. 05

   What's next

   Making Android Auto fill the truck's tall 12-inch portrait screen, and finding a dependable way to reach the Pi while parked away from home wifi.