Major League Hacking

Major League Hacking MLH is a 500k+ global member community empowering the next generation of developers to learn through and the .

08/18/2026

A hardware crypto wallet usually costs around $150. This team built one on an Arduino Uno in 36 hours.

Meet cryptX, a hackathon project from Rahul Balasubramanian, Navin Narayanan, Sakthi Sadayappan, and Akshat Goyal. The name comes from a cryptex, the portable puzzle box built to guard secrets, which is exactly what this device does for your Solana keys.

The team reverse-engineered the Ledger Nano S to understand how private keys stay isolated, then rebuilt that trust boundary themselves. Users create a 6-digit PIN, back up a BIP-39 seed phrase displayed word by word on a small LCD, and physically confirm every signing request with two buttons before a transaction ever reaches the chain. Challenge-response authentication using SipHash-2-4 ties the device identity to the physical silicon it runs on.

The part that makes it feel like a product, though, is the security layer. A Google Gemini agent monitors every connection, PIN attempt, and transaction event, turning raw device logs into real risk assessments. Repeated auth failures or a mid-transaction disconnect gets the device blacklisted automatically. All of it surfaces in a live Next.js dashboard alongside on-chain balances and transaction history.

See the full build at the link in our Instastory đź”—

08/11/2026

Most learning tools are built for a hearing kid with two hands on a keyboard, then “made accessible” with a captions toggle at the end.

V↑ starts from the opposite direction. Built by Ronald Li, Kenny Nguyen, and Sai Wong at HackDavis 2026 (Winner, Most Creative Hack), it is a conversational learning platform where the very first choice a student makes is how they want to answer: speak it, fingerspell it on camera, or photograph the work they did on paper.

Pick a character (Bear, Fox, Robot, or Cat) and they follow you into three 3D worlds: a science maze, a math boss battle, and a reading trail through the forest. Each one has its own ElevenLabs voice and personality, always-on captions, and a warm “good try, let’s look at that again” instead of a red X.

Under the hood: PyTorch models behind MediaPipe Hands read fingerspelled letters from we**am frames, Gemini Agentic Vision marks up handwritten math like a teacher and circles the first mistake, and students can upload their own lecture video to have the entire game rebuild around it.

What stands out most is the restraint. The team calls the we**am feature fingerspelling input rather than “ASL support,” because that is precisely what it is, and they added a confirmation step before grading handwriting so a misread digit never tells a struggling kid they are wrong.

Accessibility is not a feature you bolt on at the end. Nice work to the team for building like they know it.

Check out the build at the link in our Instastory đź”—l

08/06/2026

Your doorbell rings. What if it could also decide whether to be friendly or firm?

There are over a million burglaries reported in the U.S. each year, and most security systems only tell you about it after the fact. Check out Doortective, a hackathon project by Tirth Patel, .avag, Quan, and Patrick Dodds that replaces passive recording with a front door that actually responds.

A Raspberry Pi handles live camera and mic input while OpenCV and YOLOv8n detect and recognize visitors in real time. The Gemini API reads the full picture (who’s there, what they’re saying, how they’re behaving) and classifies the moment as normal, suspicious, or threatening. ElevenLabs then voices the reply in multiple languages with a tone to match, and Twilio calls the homeowner when a situation escalates. All of it in roughly one second, end to end.

Read more about their build at the link in our Instastory

08/04/2026

Attackers already have agent intelligence. Most security tooling doesn’t.

Check out Codebreaker, a hackathon project built to level that playing field. Built by Jay Wu and Kevin Wu, Codebreaker is an end-to-end platform that finds, proves, and fixes software vulnerabilities alongside security teams.

It starts with ECVE-Bench, a CVE benchmark grounded in the GitHub Advisory Database that spans the MITRE CWE Top 25 across every major ecosystem, with each task graded at four difficulty layers. From there, a security-specialized agent harness gives frontier models the tools and instincts of a vulnerability researcher instead of a generalist engineer. And when something surfaces, Devin reproduces it, writes a regression test, generates a patch, and opens a pull request — not a dashboard alert, but a merge-ready PR that drops straight into workflows teams already use.

The findings are worth sitting with: frontier models are still limited here, with GPT-5 averaging 24% per task. But one model climbed from 35% to 47% simply by being embedded in the harness, a reminder that tooling and context are a real lever on capability, not a wrapper around it.

Read more about the build at the link in our Instastory.

07/30/2026

Years of vocabulary drills won’t save you when a barista asks a follow-up question.

That gap between studying a language and speaking one is exactly what LinguaScene tackles: a hackathon project by Dhvani Gupta, Grace Li, and Natalie Ngo that turns practice into actual conversation.

Pick a scene: a doctor’s consultation in Spanish, a café order in French, a job interview in Mandarin. An AI character responds naturally in your target language while you talk your way toward a goal, using voice input and output rather than a keyboard. The Gemini API keeps the character consistent across the whole exchange, slips in corrections without breaking immersion, and closes the session with a grammar breakdown and one specific thing to improve.

Nine scenes ready to go, plus the ability to build your own. All designed, built, and deployed by a team of four in 48 hours.

Read more about their build here: https://devpost.com/software/linguascene

07/23/2026

Budgeting shouldn’t feel like a guessing game.

Meet Budgeteer, a hackathon project by Andrew Sunarto, Leyna Nguyen, and Ashley Lin built to act as your personal AI financial coach, not just another expense tracker.

Budgeteer brings your statement activity, accounts, loans, and budgets into one clear view, then uses a lightweight decision intelligence model to suggest smart budget actions tailored to you. Powered by Llama 3, it goes further than tracking by answering your financial questions in plain language so you actually understand where your money is going, not just where it went.

Read more about their build in our insta story đź”—

07/21/2026

One day until DevRelCon NYC kicks off at Industry in Brooklyn!

Let us give you a tour of this wonderfully eclectic space. We can’t wait to see you tomorrow. 🌇

Don’t have a ticket yet? You can still grab one at the link in our bio 🔗

07/21/2026

What if one photo of what’s in your fridge could hand you back a full recipe?

That’s the idea behind Dachef, a hackathon project by Saphal Poudyal rethinking how we decide what to cook.

Point your camera at your fridge or pantry, and Dachef scans the shot to identify your ingredients precisely enough to distinguish a green apple from a red one. From there, it uses Grounded Gemini 2.5 Pro to pull recipes centered on what you’ve actually got, so you’re not left stuck mid-recipe over one missing spice. It also calculates the carbon footprint of each ingredient, and through the Knot API, can suggest greener homemade alternatives based on your food delivery history.

A small but clever example of how AI can help in everyday tasks like figuring out what’s for dinner.

Check out the build: https://lnkd.in/e2-N_apr

07/16/2026

Imagine an AI companion that can speak for you when you can’t.

Check out Hear2Help, a hackathon project by Saloni Shah and Arya Gijare designed to give deaf, non-verbal, and panicked users a fast, voice-free way to call for help.

Hear2Help works in just a few taps—triple-tap to trigger SOS mode, and the app uses ’ real-time text-to-speech to communicate on your behalf. It doesn’t stop there: the tool leverages the Gemini API to answer emergency questions, pulls in geolocation to show the nearest hospital or police station, and even includes a Calm Mode with breathing exercises for panic attacks.

Read more about their build here: https://lnkd.in/ewDTGhNy

07/14/2026

Can you tell a human’s doodle from an AI’s?

That’s the question at the heart of TuringSketch, an incredible hackathon project by Jun Bin Cheng and Larry Chen that reimagines the classic Turing Test as a competitive drawing game inspired by skribbl.io.

Here’s how it works: each player gets 30 seconds to sketch a prompt while an AI generates its own version of the same image. When the timer runs out, players are shown both drawings side by side and must guess which was human-made. Jun Bin built the backend database with Azure SQL Server and designed the frontend in Streamlit, while Larry handled the AI image generation and tied the frontend and backend together. They even fine-tuned prompts so the AI would produce “MS Paint style” sketches, making the guessing genuinely challenging.

Projects like this highlight the power of builder culture—taking a foundational concept in AI and turning it into something playful, competitive, and thought-provoking!

Read more about their build here: https://devpost.com/software/idkkkkkkk

Address

Atlanta, GA

Telephone

(404) 620-6990

Website

Alerts

Be the first to know and let us send you an email when Major League Hacking posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share

Category