Codex RayBan-Meta Tutor - An English tutor that lives in your kid's glasses.

IMPORTANT: when playing the video, please turn up the volume to MAX to understand the voice output from the RayBan glasses.

The problem

Kids learning English as a second language in school get stuck waiting for a parent to read over their shoulder. Tutoring apps force typing and screen-switching — friction that breaks the flow.

The solution

Codex Tutor turns your Ray-Ban Meta glasses (with iPhone fallback) into a hands-free English tutor. The kid keeps looking at his paper homework, points at things, asks out loud. The glasses answer by voice. Corrections pop up on the phone only when needed.

Three core interactions

• Point at a word → "How do you pronounce this?" → spoken answer.
• Point at a phrase → "Is this correct?" → yes/no by voice; if no, correction on phone.
• Look at the page → "Any mistakes?" → yes/no by voice; if yes, corrections on phone.

How it works

Built on an existing dual-input app (phone camera+speakers or glasses). The RayBan glasses provide a (low-res) live-stream of video and audio to the Google Gemini Live endpoint. Gemini detects when the student asks a question (optionally pointing at a sheet of paper) and triggers a function call, that takes a high-res photo using the RayBan glasses and send this photo together with the question / input prompt to a VLM for processing. The VLM processes the students question using the image and prompt and returns a response that is converted to speech by Gemini Live. The audio is then output using the RayBan glasses.

Stack

iPhone Ray-Ban Meta glasses · phone app (live + corrections view) · VLM for pointing and for grammar · TTS for voice.

IMPORTANT: when playing the video, please turn up the volume to MAX to understand the voice output from the RayBan glasses.

When coding with agents, we typically think in linear threads, but really we should be thinking in terms of trees. If an agent presents you with options, you can choose one or open three tabs. Neither is a good option, and that’s what we’re solving!
We built a user interface for agentic coding, that reframes agents and their context as what they actually are: graphs. With the option to trivially fork off multiple agents in parallel, at ANY point.
Our system has a full container based orchestrator that manages filesystem and context. What’s more, we added the option to merge diverging branches of agents and their context into each other, so you can work with the context that’s right for your next prompt. And all of that while respecting caches, to make sure not a single token is wasted.

Our demo is fully working, no mocks here ;) We’re excited to present it in more detail!

We built is a live security benchmark for LLMs that evaluates models in real CTF environments rather than on static datasets or artificial evals. Models solve actual CTF challenges [Usually 4-5 per weekend, we select a good one], submit real flags, and compete on a shared real-time leaderboard, making it easy to see which model performs best under realistic offensive security conditions.

CTFs can be found here: https://ctftime.org/

The EU has €95 billion in grants. Their search API is broken - filters don't work, results are unsorted. We built a tool that indexes live grants, matches them to your company using OpenAI embeddings and GPT scoring, and generates application briefs. Paste a description, paste a URL, or just type a company name.

Catch your AI slop before your customers do!

We deploy six opinionated, AI Agents that test your Website with agent-browser and PI Agent.

Each Persona has a real Person behind it, for example Gergely Orosz is the inspiration for the Senior Dev POV, so we actually have a person that is in the LLM's training data.

You get feedback from 6 different points of view.