TalkBlock | Devpost

Inspiration

We love hearing innovative hackathon pitches. We don't like spending valuable time on respective buzzword slop. This weekend, we wanted a way to see the cool projects people made without the slop. Then we realized this goes well beyond just project pitches.

What it does

TalkBlock is inspired by AdBlock but for real life. It lets you experience the things you enjoy, and helps you avoid the things you don't.

You can have it running in the background of your phone and it uses your microphone to analyze the context of your conversations. When it detects you are in a situation that you don't want to be in, it creates a convenient excuse for you to leave.

Some excuses include an angry call from your girlfriend/mom/boss, or for dire situations, a fake tornado warning.

Users can set their preferences for what counts as a nuisance. They can define specific keywords that signal to get them out of a conversation, or they can describe a non-ideal situation and use our real time semantic context detection system built on OpenAI models to detect certain scenarios regardless of how they are worded.

It can free you from real life advertisers trying to sell you things or the other nuisances in life such as a work meeting running overtime, or your mom trying to make you do chores.

How we used the OpenAI API and Codex

OpenAI API Our core system, the semantic context detection, is built around the OpenAI Chat Completions API. This system is capable of performing reasoning on a constant stream of speech to detect any scenario in real time, regardless of how it is worded. In a world where systems are becoming more aware for personal intelligence, we focused much of our time on finding way to transform this API from a single response call into an ongoing constant analysis tool. We achieved this by implementing the quick and efficient responses GPT5-mini along with a short, custom implementation of a rolling context window. This allowed us to prompt engineer the advanced reasoning process that is enabled by large language models on a constant feed of text without hitting rate throttling limits. The result showed strong accuracy in most test cases ranging across very different targeted scenarios.

Codex For many of us, this was our first time using Codex and we learned how to use it to it's maximum. We stressed it to its maximum usage credits regularly for code generation and debugging. This allowed us to write more code and solve bugs that we many not have otherwise been solvable in 36 hours. Even outside of programming, Codex was incredibly valuable across the board for creative tasks such as ideation and iterative feedback. As we developed our idea, we were able to create a continuous feedback loop to expand and refine it, and as it came together, Codex helped us brainstorm names as well. Our small hackathon team suddenly had the efforts of an entire company. Because of all this, we were able to create the most advanced AI perception tool we have ever created.

How we built it

Major Components & Tools

Frontend Mobile App: Expo, React Native, custom Kotlin Android modules (for background audio), and AsyncStorage (for local data).
Speech Recognition: Google Cloud Speech-to-Text.
Backend Server: Express.js (Node.js) for secure API routing and key management.
AI & Semantic Analysis: OpenAI API (GPT-5-mini) for real-time context detection.
Telephony & Triggers: Twilio API for simulated phone calls.
Testing & Tooling: Vitest (API logic), Supertest (Express routes), JUnit (Kotlin native modules), and concurrently (unified dev environment).
Audio source: Other amazing hackers, mentors, and sponsors that we had so much fun making scripts and recording "Phone calls" with!

Process Flow Mic (Google Cloud Speech-to-Text) → Rolling Transcripts → POST /classify (OpenAI GPT-5-mini) → Context Match Detected → POST /call (Twilio) OR Tornado Alert Modal

Challenges we ran into

Finding the ideal prompt heading context to accurately predict a wide range of scenarios with the limited text input needed for fast responses.
Making everything seamless. It was surprisingly difficult to make sure all calls and alerts were indistinguishable from the real thing while keeping the app discrete.
The Twilio phone call API was a battle against rate limits and were were concerned that poor reception could ruin our demo. Lots of consideration was put into reliability and contingency.

Accomplishments that we're proud of

We believe this is the most reliable custom real time audio perception we have ever seen from a hackathon, which is really rewarding. We think we made something unique that we have not seen elsewhere.
We have surpassed the days of making the "Hackathon Special". While we took on a lot, we managed to make every button and feature work as advertised, which is rare in general, not just for a complex project like this.

What we learned

New APIs that we haven't used before.
This was our first time with Codex - very cool.
Integration hell is a real place. 80% done means 20%.
You can get some pretty cool phone numbers if you get one from Nunavut.

What's next for TalkBlock?

Dynamic Conversation - Implement speech synthesis like ElevenLabs to create phone call excuses that adapt to the user's current scenario. In the event that someone nearby Isn't convinced, it can keep gaslighting them.
iPhone deployment - we designed the architecture to be cross platform compatible but have not fully implement it on iPhone.
Publishing on public app stores
Working closer with AI providers to ensure all data is retained only for the purpose of context detection and is never saved or used for training.