Interval

A continuous, source-backed personal intelligence channel with live voice steering. It tracks changes across your interests, remembers what you’ve heard, and replans what comes next in real time.

Comment

The inspiration

The world is moving fast, too fast even. I had a really tough time keeping up with all the latest developments. New AI models, new product announcements, economical shifts, political developments.

Intelligence has always been critical, make or break, even. So I thought, what if, intelligence could be automated?

The project

Interval collects the latest developments according to your interests, and streams it to you as a live voice intelligence channel.

Turn any downtime into an opportunity for you to stay up to date with the world. Interrupt or speak to Interval any time for deeper dives, questions, and steering.

The process

Interval is built entirely using OpenAI Codex, powered by the OpenAI API.

Codex with GPT-5.6 Sol handled everything from discussion to build.

The idea started in the inception thread with GPT-5.6 Sol, where the vision and implementation was refined. When the product solidified, an LLM wiki scaffold was established for enabling shared knowledge and tracking progress between threads.

The engineering thread was mainly where Interval was built, GPT-5.6 Sol on Ultra mode handled everything from backend, frontend, API access, design, and live tested the app using computer use.

More threads were created and driven for further refinement of the app, as well as the buildout of the landing page and the iOS app.

The app itself is powered by gpt-5.6-sol and gpt-5.6-luna, narration, voice inputs, and live responses powered by gpt-realtime-2.1.

The challenges

Layering the various models to produce curated content, while being steerable live, was one of the biggest challenges of this project.

Interval initially started out using gpt-5.6-sol for content curation, tts-1 for speech generation, and gpt-realtime-whisper for voice input. This wasn't entirely ideal however, as there was a pause where the voice input directives had to be processed by gpt-5.6 first before returning back as speech output, which felt unnatural and resulted in bad UX.

After much back and forth with GPT-5.6 Sol on Codex and numerous iterations, we've opted to use gpt-5.6-luna for faster curation, and unified the voice input and output using gpt-realtime-2.1, which allowed not only consistent voice throughout but also very low latency, tool call backed ad-hoc web research that allowed topics to be explored and steered live.

Future developments

Interval will be launched as a real product on desktop, iOS, watchOS, and Android. Looking forward to updating Interval to use the GPT-live-1 model once it’s available on the API.

Built With

Share this project:

Updates