[01 // Transmission]
Building an AI That Encourages You to Look Away
What if an AI application wasn't designed to keep you on the screen, but to help you leave it?
Most applications compete for our attention. AI applications often take this even further, offering endless conversations, instant answers, and another reason to keep looking at a screen.
While working on the first week of Hacktoberfest 2026's Open-Source AI Challenge, I wanted to explore a different idea.
What if AI could help us pay more attention to the world instead of demanding more of our attention?
That question led me to build Observation Coach.
The Idea: Observe First, Ask AI Later
Observation Coach is an open-source AI application designed to help people practise noticing the details around them.
The idea is deliberately simple. Instead of asking AI to identify everything through a camera or explain the world immediately, the application encourages users to step away from their devices, observe something themselves, and return to reflect on what they noticed.
The AI acts as a coach rather than an answer machine.
How It Works
The experience follows a short, guided cycle:
- Choose an exercise. Select an observation session lasting 5, 10, 20, or 30 minutes.
- Step away from the screen. Go out and pay attention to your surroundings.
- Record what you noticed. Return and describe an actual observation in your own words.
- Reflect with AI. The coach asks a follow-up question based on your description.
- Keep learning. Receive an adaptive exercise to help develop your observation skills over time.
The focus is on attention, description, comparison, change, inference, and deeper observation.
The goal isn't to make every ordinary moment dependent on AI. It's to help people become more curious and observant on their own.
Building the Application
I built Observation Coach with a focus on modular architecture and flexibility in how AI inference is handled.
The technology stack includes:
- Next.js App Router for the application.
- TypeScript for type-safe development.
- Tailwind CSS for the interface.
- MongoDB for user accounts, observation sessions, saved observations, and progress.
- Qwen 3B Instruct in GGUF format, run locally through
llama.cpp. - Qwen 27B Instruct through Groq for the hosted inference configuration.
I also implemented structured response validation so model outputs can be checked before being used by the application.
One interesting challenge was designing the AI's role. The coach shouldn't immediately turn a simple observation into a lengthy explanation. It should encourage the user to think a little further, ask a useful question, and suggest an appropriate next exercise.
That makes the interaction less about generating impressive responses and more about supporting a learning process.
Why Open-Weight AI Matters Here
I wanted the application to offer flexibility instead of being permanently tied to a single hosted AI provider.
With a local model running through llama.cpp, developers can experiment with inference on their own hardware. A hosted, OpenAI-compatible API provides another option when local hardware isn't sufficient.
This architecture creates a useful distinction:
- Local inference: offers greater control over where personal observations are processed.
- Hosted inference: offers an alternative when running a model locally isn't practical.
- Provider flexibility: makes it possible to experiment with different models without redesigning the entire application.
The project also includes a guided demo mode so people can explore the observation workflow without needing a working AI API key.
Local inference can support an offline setup when the application, database, model, and required dependencies are all available locally. The hosted configuration, of course, depends on network access.
The Design Decision I Care About Most
The most important part of Observation Coach isn't the model size or the technology stack.
It's the decision to keep the AI interaction short.
The application gives the user an exercise, steps aside, and waits for them to return with something they actually noticed.
That is a different approach to designing an AI experience. Rather than maximising the time spent interacting with the assistant, the application tries to make the interaction useful enough that the user can return to the real world.
I find that idea worth exploring: technology doesn't always need to occupy more of our time to be valuable.
What I Took Away From This Project
Building Observation Coach gave me an opportunity to explore more than model integration.
It involved thinking about the product's interaction flow, structuring model responses, separating local and hosted inference, and designing an application around a goal that isn't simply more screen time.
It also reminded me that a useful AI product begins with a clear problem and a thoughtful workflow, not just a model API.
There is still room to improve the experience, but the core idea is something I want to continue exploring.
Try It Yourself
If you'd like to explore the project, here are the links:
- Live demo: https://observation-coach.vercel.app
- Source code: https://github.com/devrittik/observation-coach
- Product walkthrough: https://youtu.be/KArn5bRtpeI
- Code walkthrough: https://youtu.be/vQHZ7amHMTo
Observation Coach was built for Hacktoberfest 2026 — Open-Source AI Challenge, Week 1: Touch Grass.
The idea I'd like to keep exploring is simple: what if the best AI interaction is the one that helps you confidently put your device away?