This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
What I Built
I built GrassWhisper, a local AI notification gatekeeper designed to help people spend time outside without being constantly pulled back into their phones.
The idea is simple: when you go for a walk, your phone still receives notifications, but you should not have to look at the screen to decide which ones matter.
GrassWhisper classifies notifications into three levels:
- 🔴 Critical — speak immediately
- 🟠 Important — give a brief interruption
- 🟢 Queued — stay silent and save it for later
It also goes beyond one-notification-at-a-time triage.
GrassWhisper can:
- summarize bursts of messages from the same conversation instead of announcing every notification separately
- detect repeated contact such as missed calls followed by messages and escalate the alert
- merge notifications from different apps when they refer to the same event
- use Privacy Mode to avoid speaking private message contents in public
- provide Full Readout Mode when the user explicitly wants the message spoken
- track the No Screen Challenge, including walk duration, screen time, notifications received, notifications allowed through, and notifications suppressed
- use motion-sensor information in supported mobile browsers to help detect walking automatically, while still keeping a manual Walk Mode toggle
The goal is not to help you manage more notifications.
The goal is to help you stop looking at notifications altogether.
Video Demo
Live / video demo: https://youtu.be/R57HdKb-YjQ?si=QHnu9WVGNwFiys3I
The demo shows GrassWhisper entering Walk Mode, receiving different types of notifications, filtering them locally, speaking only the notifications that need attention, and producing a post-walk No Screen Challenge summary.
Code
GitHub repository: https://github.com/ABHISIKTAcommits/GrassWhisper
How I Built It
GrassWhisper is built around local open-source AI inference rather than sending notification content to a cloud AI service.
The main components are:
- Python + FastAPI for the local backend
- Ollama for local model inference
- an open-weight local language model for notification triage
- HTML/CSS/JavaScript for the interface
- the browser's Web Speech API for offline spoken alerts
- browser motion-sensor APIs for optional walking detection on supported mobile browsers
The AI is given the notification context and the user's Walk Mode filter, then produces a structured urgency decision.
The application does not simply trust one model response. Deterministic safety and priority rules are applied around the model so that things such as OTPs, emergencies, missed calls, schedule changes, deadlines, and direct requests are handled consistently.
The notification pipeline is therefore:
Notification → Local AI + rules → Urgency level → Aggregation → Privacy-safe audio → Post-walk digest
Why Does Open Innovation Matter?
Privacy is one of the main reasons this project benefits from open AI.
Notifications can contain extremely personal information: banking messages, family conversations, work incidents, appointments, authentication codes, and private requests.
Sending that information to a closed cloud AI service introduces another system that has to receive and process it.
With local inference, GrassWhisper can keep the notification-processing workflow on the user's own machine instead.
Open AI also gives the project more control over how notification importance is defined. The model can be swapped, the prompts can be changed, the rules can be inspected, and the behavior can be adapted without depending entirely on a proprietary API.
For a project whose purpose is to help people put their phones away, keeping the intelligence as local and controllable as possible is especially important.
Prize Categories
Overall — Hacktoberfest Open-Source AI Challenge: Week 1 — Touch Grass
GrassWhisper uses open-source/open-weight AI locally through Ollama, with the goal of reducing screen dependence and helping users spend more time outside.
Top comments (0)