https://github.com/jinkang1st/Hackathon2026.git
Inspiration
Digital life makes acting on impulse effortless. Buying something expensive takes one click. Sending a message in frustration takes one second. Regret often arrives just a little later.
We built Pause — Your Personal AI Firewall around a simple question: what if technology gave us a moment to reconsider?
Pause creates that moment, helping people make decisions that reflect their own intentions, boundaries, and priorities.
What it does
Pause brings timely reminders into everyday browsing and communication. Our prototype focuses on two scenarios:
- Shopping Guard: A Chrome extension detects product prices and reminds users when an item meets or exceeds their chosen spending threshold. It works with local GBP price detection and supports optional AI analysis to interpret product information.
- Message Guard: A native macOS app checks drafts in applications the user explicitly enables. When it detects potentially harmful language, AI evaluates the context and can suggest a clearer, less confrontational alternative.
The current prototype provides voluntary reminders while browsing and typing. Users decide whether to continue, reconsider, or use a suggestion.
How we built it
We built three connected components:
- A Chrome Manifest V3 extension using JavaScript for page analysis, spending settings, and reminders.
- A native Swift macOS app for checking supported text fields in approved applications and displaying suggestions.
- A Python and FastAPI backend, using the OpenAI Python SDK to connect to models hosted through Nebius Token Factory.
We combined lightweight local checks with AI assessment. Shopping reminders can work without a model connection, while message drafts reach AI assessment only after matching the local screening rules.
The backend also combines structured model responses with deterministic checks, keeping explicit user rules separate from model judgment.
Challenges we ran into
Context was one of our biggest challenges. An offensive word in a quotation should be treated differently from a personal attack. Similarly, a crossed-out price or monthly instalment should not be confused with a product’s total price.
Timing also mattered. Users keep typing and switch applications while AI requests are running. We needed to discard outdated results so reminders stayed relevant.
Another challenge was making interventions useful without becoming intrusive. We focused on clear explanations, limited context collection, explicit application permissions, and controls that keep the user involved.
Accomplishments that we're proud of
We built a working browser extension, native macOS companion, and shared AI backend within the hackathon.
We also validated the message workflow with real model inference: a physical keyboard event triggered local screening, contextual assessment, and a native reminder in TextEdit.
We are especially proud that shopping reminders remain useful without AI connectivity, and that message suggestions require a deliberate user action before being copied.
What we learned
Useful AI assistance depends on more than generating a good answer. It needs the right timing, enough context, reliable fallback behavior, and a clear role in the user’s decision.
We learned that a small, well-explained intervention can be valuable. Sometimes the most helpful thing technology can offer is a moment to think.
What's next for Pause — Your Personal AI Firewall
Next, we want to improve price detection, support more currencies and applications, and evaluate reminder quality across more real-world situations.
We also plan to explore application-specific checks before sending messages, optional cooling-off periods for purchases, and safeguards around sensitive information sharing and important file deletion.
Our goal is to make Pause a personal safety layer that helps people act with intention, one decision at a time.
Log in or sign up for Devpost to join the conversation.