Inspiration
GuardianNode was inspired by an interview featuring cybersecurity researcher Ryan M. Montgomery, where he discussed the real-world dangers children can encounter online.
The interview made me worry that I had no real visibility to my kids activity online. Parents are expected to protect their children across games, social platforms, browsers, messaging apps, and rapidly changing online communities, but most existing tools offer only blunt blocking controls or invasive cloud surveillance. I put my kids on a vLAN that filters and warns me but I don't think it was enough.
Parents often receive an alert without enough context to know what it means. A concerning phrase might indicate grooming, coercion, bullying, self-harm risk, or exploitation—but it might also come from a game, a joke, school research, a news story, or a conversation being quoted out of context.
I wanted to build something that could help parents identify possible risks without continuously sending a child’s private activity to someone else’s cloud. I wanted it to also be better then just blocking a website or filtering sites. The tools we have now with Codex and ChatGPT can make anyone a developer. I have been interested in technology for my whole life but I chose to spend most of my life in management and tech is my hobby. I felt I could use my skills to make an app that would help others. Without modern AI I would have spent years getting close to an app like this.
These concerns became GuardianNode.
What GuardianNode does
GuardianNode is a local-first family safety platform for parent-managed devices.
It can analyze activity locally, identify events that may deserve parental attention, preserve relevant evidence, and present those events through a parent dashboard. The goal is not to automatically judge or punish a child. The goal is to give the parent enough information to understand what may have happened and decide what to do next. It can see what a typical router filter will miss.
GuardianNode keeps continuous monitoring and original evidence local whenever possible.
During OpenAI Build Week, I am extending the existing GuardianNode alpha with a new feature called Guardian Review. I felt this would be helpful to parents that need guidance on speaking with their children.
Guardian Review
Guardian Review uses GPT-5.6 to help a parent reason through an ambiguous safety alert.
When a parent opens an incident, GuardianNode first explains why its local detection system flagged the activity. The parent can then optionally request a Guardian Review. I still wanted to focus on local first and privacy. Before the request is sent, GuardianNode minimizes the incident context, redacts unnecessary identifying information, and shows the parent exactly what will be shared.
GPT-5.6 then provides a structured review that can include:
- What was directly observed
- Why the incident may be concerning
- Possible benign explanations
- Important missing context
- The model’s confidence and limitations
- Questions the parent may need to answer
- Proportionate immediate and follow-up actions
- Indicators that may justify escalation
- Guidance for approaching the child calmly
- Suggested opening language and questions
- Accusatory or counterproductive approaches to avoid
Guardian Review is designed to help the parent understand both the potential risk and the uncertainty surrounding it.
It does not automatically restrict the child, contact another person, report an incident, diagnose a condition, or make the final decision. Sometimes letting the kids experience something then good parenting is the best medicine.
The AI advises. The parent decides.
How I built it
GuardianNode entered Build Week as a functioning alpha with a Windows agent, local processing components, backend services, a parent dashboard, evidence handling, detection workflows, and an installation path.
The core GuardianNode platform was built primarily with OpenAI Codex. Claude assisted with portions of the pre-existing web interface.
For Build Week, I am using a dedicated Codex development session powered by GPT-5.6 to design, implement, test, and integrate Guardian Review.
Codex has helped with absolutely everything. Codex has written most all of the code, the images, the APIs, the backend and frontend, and essentially the entire codebase. I gave insights, tested, helped guide the development, gave the ideas, and set the guardrails.
GPT-5.6 is used in two distinct ways.
First, it powers the Codex development process used to build and harden the feature.
Second, it is used inside Guardian Review at runtime to interpret context, identify uncertainty, consider competing explanations, and turn a technical alert into practical guidance a parent can understand.
Privacy by design
Family safety software handles unusually sensitive information, so privacy cannot be an afterthought.
GuardianNode is designed around several principles, keep the data local first, give the kids some privacy, and help parents know if something concerning is happening before its too late.
GuardianNode is not intended to replace parental judgment, qualified mental-health care, emergency services, or law enforcement. It is intended to help parents notice possible problems sooner and respond with more context and care.
Challenges
The largest challenge was not simply connecting an AI model to an alert.
The difficult part was designing a workflow that remains useful without giving the model too much authority.
A weak implementation could easily become alarmist, label incomplete evidence as fact, recommend disproportionate action, or encourage a parent to confront a child based on an incorrect interpretation.
Guardian Review therefore has to distinguish between observations and inferences, communicate uncertainty clearly, identify missing information, and consider realistic benign explanations.
Another challenge was maintaining GuardianNode’s local-first privacy model while still gaining the benefit of GPT-5.6. That required a deliberate minimization, redaction, preview, and consent process instead of automatically uploading raw screenshots or complete conversations.
I also feel its important for the kids to have some autonomy. I don't want the app to be a "big brother." I want my kids to be able to speak to their friends, joke, have relationships, and enjoy the time online. I just want to be aware of the external influence on them. The internet has always been a dangerous place but things were different when I was a kid.
What I learned
Building Guardian Review reinforced that responsible AI design is not only about adding warnings to the interface.
The system architecture must reinforce the intended decision boundary.
The model should not receive more data than it needs. It should not control enforcement. Its output should be structured, reviewable, and attributable to a specific prompt and schema version. Parents should be able to see uncertainty rather than receiving a confident paragraph that appears authoritative.
I also learned how valuable Codex can be when working across an established product rather than a small isolated demo. It helped trace existing workflows, connect new functionality across the agent, backend, and dashboard, create tests, and continuously validate that new work did not break the existing platform.
Accomplishments
The most important accomplishment is turning Guardian Review into a complete product workflow rather than a generic chatbot. A product that is free, helps parents be there, and gives kids a chance to make mistakes and have someone to guide them. If my app can save a single kid from harming themselves, being harmed by someone else, or help a parent get involved before its too late, then I have accomplished my goal.
A parent can move from:
local detection → incident evidence → privacy preview → GPT-5.6 review → communication guidance → human decision
What is next
The immediate goal is a GuardianNode beta that technical parents and evaluators can install and test safely using synthetic scenarios. I would like to extend this to Android and iOS at some point.
Future work includes:
- Broader evaluation across different age groups and family situations
- Improved local personalization without cloud profiling
- More transparent confidence calibration
- Expanded accessibility
- Additional parent feedback and follow-up tools
- Safer remote notifications
- Independent privacy and security review
- More robust Windows packaging and update support
- Carefully designed support for additional operating systems
GuardianNode began with a simple goal: help parents recognize danger without creating another source of danger through excessive surveillance or automated judgment.
Guardian Review extends that goal by helping parents understand not only what may have happened, but also how to respond in a way that protects both safety and trust.
Built With
- computer-vision
- family
- fastapi
- gpt-5.6
- local-ai
- ocr
- ollama
- openai-api
- openai-codex
- privacy
- python
- react
- responses-api
- structured-outputs
- typescript
- windows
Log in or sign up for Devpost to join the conversation.