• 

• 

• 

• 

• 

• 

• 

🛡️ MindSteady

The Real-Time Runtime Shield for Children's Video Safety

Parental controls today only block explicit words and themes. They do nothing about the hyper-stimulating video pacing that overwhelms a developing brain.

MindSteady is a real-time, AI-powered runtime shield. It intercepts video as it plays and gently throttles hyper-fast scene cuts, dampens harsh contrast, and stabilizes playback — before sensory overload happens.

🚨 MindSteady doesn't just flag content. It intervenes — live, on the page, while the child is watching.

---

📖 Table of Contents

-⚡ TL;DR

• 💡 Inspiration
• 🔥 The Problem: The "Passive Filter" Flaw
• 💡 The Solution: An Active Runtime Shield
• 🎯 Target Users
• 🧩 Key Pillars
• 🏗️ System Architecture
• 🔄 How It Works: The Closed-Loop Pipeline
• 🤖 The Shield Agent: Multi-Agent AI
• 🧠 Grounded in Research & Evidence
• 📚 What We Learned
• 🧗 Challenges
• 📊 Impact Model
• 🔐 Privacy by Design
• 📈 Results
• 🗺️ Roadmap & Future Scope
• 🏆 Why MindSteady
• 👤 Author

---

⚡ TL;DR

One line: MindSteady is a Chrome extension + AI backend that intervenes in hyper-stimulating video in real time — slowing playback, calming visuals, and warning parents — instead of merely blocking content after the damage is done . 🗓️ Build window: October 13–14, 2026

Question	Answer
What is it?	A real-time runtime shield that modulates video playback to protect developing brains.
Who is it for?	Parents, caregivers, and educators of children aged 0–8.
What's the AI?	A Google ADK multi-agent pipeline (3 parallel specialists + a synthesis agent) that scores video pacing, visual intensity, and content safety from sampled frames.
Why is it different?	Everyone else filters. MindSteady modulates — the intervention happens live, not in a blocklist.
Is it built?	Yes — a working MVP across three integrated components (extension, API, agents).

---

💡 Inspiration

It started with a small, familiar family struggle — and ended with a public-health concern.

One of us came across an Instagram video explaining how children's content is deliberately engineered to capture attention and create dependency — to the point where many kids cannot perform routine tasks, like eating a meal, without being handed a phone. That resonated immediately with our own observations of younger children in our families, who insist on screen time whenever food is involved.

What felt like a minor parenting annoyance suddenly connected to something much larger: if this pattern continues unchecked, it could have serious long-term effects on children's development.

🔍 What We Found	📌 The Implication
Popular shows are reportedly engineered using child testing to maximize "stickiness" rather than educational value.	Developing brains are exposed to engineered addiction, not enrichment.
Modern children's content often relies on sensory overload — scene changes every 1–3 seconds, extreme color saturation, and rapid stimuli beyond natural human processing capacity.	The harm is in the delivery, not the topic — invisible to every text-based filter.
A single flagship channel generates 7.8 billion monthly views and 134 million subscribers, reaching millions of infants during critical brain-development stages.	The exposure is massive and systemic, not a fringe case.
Studies link high screen exposure to structural brain changes, altered EEG patterns, elevated stress hormones, and persistent executive-function deficits that can last into school age.	The effects are measurable — this is a real, defensible problem.

And yet there is a regulatory gap. No existing technology automatically detects and moderates the visual characteristics — pace, contrast, content safety — that research has proven harmful.

Motivated by both personal experience and this growing body of evidence, we built MindSteady: a protective technology layer designed to safeguard the developing brains of infants and toddlers.

---

🔥 The Problem: The "Passive Filter" Flaw

Traditional parental controls are a locked door on a house that is already on fire.

Every major parental-control tool today runs on the same outdated model: block or allow. They scan titles, transcripts, tags, and known keywords, then decide whether a video is permitted. This leaves a massive, unaddressed gap.

❌ What Existing Tools Catch	🚨 What They Completely Miss
Explicit language in metadata	Hyper-fast scene cuts that fragment attention
Known banned keywords	Strobing & flashing that overwhelm visual processing
Category-based age ratings	Extreme saturation & contrast spikes
Static allow/deny lists	Sudden volume spikes that startle infants
Nothing about delivery	Rhythm & pacing that dysregulates developing brains

The result: a video that is correctly labeled, correctly rated, and correctly filtered can still be a sensory assault on a toddler's nervous system. Existing tools have zero runtime awareness of how content is delivered — only what it contains.

📉 The Scale of the Gap

Signal	Reality
Content moderation focus	Almost entirely on text and themes
Pacing / edit-rhythm detection	Effectively non-existent in consumer parental tools
Intervention timing	Post-hoc blocking — after exposure begins
Child-facing feedback	Surveillance dashboards, not protective modulation

---

💡 The Solution: An Active Runtime Shield

MindSteady does not just tell parents a video is fast — it steps in.

It intercepts HTML5 <video> elements in real time and dynamically applies protective measures during playback, without blocking the child's experience outright.

🎯 Intervention	⚙️ Mechanism	🧠 Why It Matters
⏱️ Playback Throttling	Dynamically reduces playback rate to 0.85x / 0.75x during detected hyper-pacing	Gives the developing brain time to process each frame
🎨 Visual Dampening	Applies calming CSS filters — brightness(0.8) contrast(0.9) saturate(0.7)	Reduces overstimulation from harsh contrast & saturation
🔇 Volume Stabilization	Lowers sudden volume spikes during intense segments	Prevents startle responses in infants
⚠️ Gentle Warnings	Non-blocking notice with a continue / skip choice	Keeps parents informed without surveillance

### 🛡️ MindSteady = a digital **seatbelt** for cognitive development *Not a locked door. A protective layer that moves with the child.*

---

🎯 Target Users

User	Pain Point	How MindSteady Helps
👨‍👩‍👧 Parents & caregivers	Can't vet every second of every video	Automatic, live protection with no manual setup
🧑‍🏫 Early-years educators	Show "safe" clips that still overstimulate	Playback modulated for the classroom age band
👶 The child	Experiences overload they can't articulate	Sensory load is reduced transparently and gently
🩺 Pediatric / OT professionals	No tool to demonstrate pacing risk	Concrete, measurable runtime intervention

Design stance: protect the child without surveilling the child. MindSteady never builds a profile, never logs viewing history, and never tracks identity.

---

🧩 Key Pillars

1️⃣ ⚡ Active Intervention

Solving the "Passive Filter" Flaw
MindSteady does not stop at flagging. It intercepts HTML5 video elements in real time to dynamically reduce playback speed, apply calming visual CSS overlays, and lower volume spikes during detected hyper-pacing segments. Real-world impact beyond simple content flagging.

2️⃣ 🤖 Google ADK Multi-Agent AI

Sophisticated, meaningful AI
Rather than a single monolithic script, MindSteady runs a parallel multi-agent pipeline built on the Google Agent Development Kit — each agent specializing in a distinct dimension of safety. Pacing · Visual Safety · Content Safety → Policy Synthesis.

3️⃣ 🧠 Research-Grounded

Real necessity, not a gimmick
Grounded in developmental research on rapid media pacing and executive-function deficits in early childhood, MindSteady targets the actual mechanism of harm. A protective layer for cognitive development.

4️⃣ 🔐 Privacy-Respecting

Protection without surveillance
MindSteady operates locally on the client layer via the browser extension. It intercepts playback without logging private viewing history, profiling the child, or tracking identity. Safety that supports independence.

---

🏗️ System Architecture

MindSteady is a monorepo built as a continuous, closed-loop feedback system across three specialized components.

┌─────────────────────────┐       ┌─────────────────────────┐       ┌─────────────────────────┐
│    Chrome Extension     │       │       Django API        │       │  Shield Agent (ADK)     │
│   (DOM Interception)    │ ────► │  (Analysis & Routing)   │ ────► │  (Multi-Agent Scoring)  │
│                         │       │                         │       │                         │
│ • Intercepts <video>    │       │ • Ingests frame data    │       │ • Pacing Agent          │
│ • Executes CSS filters  │ ◄──── │ • Dispatches policy     │ ◄──── │ • Visual Safety Agent   │
│ • Controls rate (0.75x) │       │ • Caches safety results │       │ • Content Safety Agent  │
│ • Non-blocking warning  │       │ • yt-dlp 5s sampling    │       │ • JSON Policy Synthesis │
└─────────────────────────┘       └─────────────────────────┘       └─────────────────────────┘

🔧 Component Matrix

Layer	Component	Stack	Responsibility	Status
🖥️ Client	extension/	JS · HTML · CSS · Manifest V3	Intercepts <video>, executes CSS filters, controls playback rate	✅ Working
🌐 Service	backend/	Python · Django · DRF · yt-dlp · OpenCV	Downloads a 5s sample, extracts frames, dispatches analysis, caches results	✅ MVP
🧠 Intelligence	shieldagent/	Python · Google ADK · LiteLLM	Parallel multi-agent scoring & JSON policy synthesis	✅ MVP
🎨 Brand	docs/	SVG	Logo & identity assets	✅ Complete

---

🔄 How It Works: The Closed-Loop Pipeline

[ 👶 Child / Browser ]  -- (1. Plays video) -->  [ 🖥️ Chrome Extension ]
                                                            │
                                                     (2. Intercepts <video>)
                                                            │
                                                     (3. POST video URL)
                                                            ▼
                                                     [ 🌐 Django API ]
                                                            │
                                                     (4. Fetches ~5s stream)
                                                            │
                                                     (5. Dispatches frames)
                                                            ▼
                                              [ 🤖 Shield Agent (ADK) ]
                                                            │
                                              ┌─────────────┴─────────────┐
                                              ▼                           ▼
                                      [ ⏱️ Pacing ]               [ 🎨 Visual Safety ]
                                              │                           │
                                              └─────────────┬─────────────┘
                                                            ▼
                                                   [ 🚸 Content Safety ]
                                                            │
                                                 (6. Synthesizes policy)
                                                            ▼
                                                   [ 📋 Policy Agent ]
                                                            │
                                                (7. Returns JSON layout)
                                                            ▼
                                                     [ 🌐 Django API ]
                                                            │
                                               (8. Sends action parameters)
                                                            ▼
                                              [ 🖥️ Chrome Extension ]
                                                            │
                                              (9. Throttles rate, CSS, warnings)
                                                            ▼
                                        [ Live Protective Layer Applied ]
                                        *No content blocked or blacked out*

🪜 Three-Stage Execution

Stage	What Happens
1️⃣ Interception	The Chrome Extension intercepts video elements on web pages and extracts live metadata + the page URL.
2️⃣ Analysis	The Django backend downloads a short sample (yt-dlp), extracts representative frames (OpenCV), and dispatches them to the Shield Agent pipeline.
3️⃣ Execution	The agents synthesize a structured JSON policy; the Extension applies live playback-rate throttling and visual-dampening CSS in real time.

---

🤖 The Shield Agent: Multi-Agent AI

The intelligence core is a SequentialAgent pipeline built on Google ADK: a parallel research phase feeds a synthesis phase.

               ┌───────────────────────┐
               │   🎞️ Video Frames      │
               └───────────┬───────────┘
                           │
                           ▼
               ┌───────────────────────┐
               │    ParallelAgent      │
               │    InfoGatherer       │
               └─────┬─────┬─────┬─────┘
                     │     │     │
         ┌───────────┘     │     └───────────┐
         ▼                 ▼                 ▼
 ┌───────────────┐ ┌───────────────┐ ┌───────────────┐
 │  ⏱️ Pacing     │ │ 🎨 Visual     │ │ 🚸 Content    │
 │  Agent        │ │  Safety Agent │ │  Safety Agent │
 └───────────┬───┘ └───────┬───────┘ └───────────┬───┘
             │             │                     │
             └───────────┐ │ ┌───────────────────┘
                         ▼ ▼ ▼
               ┌───────────────────────┐
               │    📋 Policy Agent    │
               │ MindSteadyCoordinator │
               └───────────┬───────────┘
                           │
                           ▼
               ┌───────────────────────┐
               │ 📦 Structured JSON    │
               │    Policy Payload     │
               └───────────────────────┘

Agent	Model	Evaluates	Output Key
⏱️ Pacing Agent playback_speed_agent	gpt-4.1-mini	Fast movements, quick cuts, rapid scene changes, information density, motion patterns that overwhelm developing visual systems	needs_slower_playback recommended_factor
🎨 Visual Safety Agent color_contrast_agent	gpt-4.1	High-contrast elements, bright / saturated colors, rapid color changes, flashing effects	needs_reduced_contrast
🚸 Content Safety Agent content_appropriation_agent	gpt-4.1-mini	Violence, frightening imagery, adult themes, scary visuals, age-inappropriate concepts	contains_inappropriate_content safety_message
📋 Policy Agent MindSteadyCoordinator	gemini-2.0-flash	Synthesizes all three telemetry streams into a single actionable recommendation	overall_recommendation

💡 Why multi-agent? Different dimensions of harm need different reasoning. Splitting pacing, visuals, and content into specialized agents yields sharper judgments than one generalist model — and the parallel architecture keeps latency low enough for real-time use.

🧬 Engineering Highlights

Technique	Where	Why It Matters
ParallelAgent fan-out	shieldagent/agent.py	Three specialists run concurrently, not in series
SequentialAgent pipeline	shieldagent/agent.py	Research phase → synthesis phase, deterministically ordered
Scene-change frame sampling	backend/ai.py	Samples every ~2s instead of every frame — cheap, representative
Multi-provider models	LiteLLM + OpenAI + Gemini	Best model per task, not one model for everything
Result caching	views.py (Django cache)	Repeat views are instant and cost-free
Graceful degradation	background.js	On API failure, fails safe rather than failing open

---

🧠 Grounded in Research & Evidence

MindSteady is not built on vibes — it targets a documented developmental mechanism. The evidence base spans neuroscience, developmental psychology, and platform-scale exposure data.

🔬 Finding	🎯 Design Response
Rapid media pacing is associated with measurable executive-function deficits in early childhood.	⏱️ The Pacing Agent actively throttles playback speed during hyper-paced segments.
Scene changes every 1–3 seconds exceed natural human processing capacity and drive sensory overload.	⏱️ Scene-change frequency is monitored and playback is slowed above a safe threshold.
Strobing, flashing, and extreme contrast overwhelm the developing visual system.	🎨 The Visual Safety Agent triggers real-time CSS dampening filters.
High screen exposure is linked to structural brain changes, altered EEG patterns, and elevated stress hormones.	🧩 The intervention targets net sensory load, not just isolated features.
Sudden volume spikes trigger startle responses in infants.	🔇 Volume stabilization lowers abrupt intensity.
Cognitive load from dense, fast-cut content degrades attention and self-regulation.	🧩 Combined interventions reduce net sensory load in real time.

Why this matters for the "regulatory gap": no existing technology automatically detects and moderates the visual characteristics (pace, contrast, content safety) that research has proven harmful. MindSteady closes that gap.

---

📚 What We Learned

Domain	Takeaway
🧠 Neuroscience foundations	How screen exposure — especially today's high-contrast, fast-paced content — shapes early cognitive development by altering brain-wave patterns, attention mechanisms, and executive function.
🎛️ Applied multimodal AI	How multimodal AI can analyze multiple video features simultaneously (in parallel) to address a real-world social problem.
🤖 AI orchestration with ADK	The practical challenges of deploying agents with the Google Agent Development Kit — including coordinating multiple agents for real-time, low-latency analysis and decision-making.
⚖️ Product judgment	How to balance protective intervention against usability, so content stays watchable rather than frustrating.

---

🧗 Challenges

Challenge	How We Approached It
⚡ Real-time performance — achieving reliable analysis without playback lag was the single biggest technical challenge.	Sample a short window (≈5s) instead of the whole video, extract a small set of representative frames, and fan out analysis to parallel agents; cache results so repeat views are instant.
📚 New toolkit learning curve — mastering Google ADK during the hackathon required rapid experimentation and adaptation.	Built the pipeline incrementally: single agent → ParallelAgent fan-out → SequentialAgent synthesis.
🎚️ Balancing intervention and usability — designing filters and playback adjustments that protect children without making content unwatchable.	Conservative, reversible interventions: gentle speed factors (0.75x–0.85x), modest CSS dampening, and a non-blocking warning with a continue/skip choice.
🌐 Cross-platform testing — ensuring consistency across video formats and platforms like YouTube and Instagram.	Target the standard HTML5 <video> element and re-scan dynamically loaded players every 5 seconds.

---

📊 Impact Model

MindSteady is designed to scale to every site with an HTML5 <video> element — that is the entire modern web.

Dimension	Today (MVP)	With Adoption
Coverage	Any page with <video> (YouTube, news, social)	Universal via browser extension
Cost per view	Near-zero (cached, short sample)	Constant, cache-amortized
Latency	Sub-second sample + parallel agents	Further reduced via edge caching
Child data collected	None	None — by design
Measurable outcome	Sensory-load reduction per session	Aggregate (non-identifying) pacing stats

🎯 Measurable Success Metrics

1. Sensory-load reduction — average pacing/contrast delta applied per flagged video.
2. Intervention precision — % of interventions judged appropriate by caregivers.
3. Time-to-protection — latency from play event to applied modulation.
4. False-positive rate — videos modulated that did not need it.

---

🔐 Privacy by Design

MindSteady is built to protect children without surveilling them.

✅ What We Do	🚫 What We Never Do
Analyze a short sample of the video on your own backend	Upload or store the child's viewing history
Keep user settings in browser-local storage	Build a profile of the child
Operate on the client layer for intervention	Track identity, location, or personal data
Communicate over HTTPS-only endpoints	Sell or share any data

"Parental tools that support safety without excessive surveillance." — aligned with the hackathon's own guidance.

---

🧪 Demo & Testing

Scenario	Expected Behavior
▶️ Safe content	Normal playback — no intervention
⚠️ Moderate content	Reduced speed + calming visual filters
🛑 Risky content	Warning dialog + pause, with continue / skip choice

# 1. Load the extension (see Quick Start)
# 2. Visit any site with video (YouTube, news, social)
# 3. Open the popup → watch live detection stats
# 4. Press F12 → follow the "MindSteady:" console trace

---

📈 Results

We tested MindSteady on a set of popular children's videos. The extension:

✅ Observed Outcome	Mechanism
Successfully adjusted playback speed when content had rapid scene changes	⏱️ Pacing Agent → runtime playbackRate control
Reduced overstimulating visual intensity through adaptive filtering	🎨 Visual Safety Agent → CSS dampening
Triggered warnings for videos flagged as developmentally inappropriate	🚸 Content Safety Agent → non-blocking continue/skip notice

These outcomes demonstrated the feasibility of real-time AI protection for infants and toddlers during digital content consumption.

---

Technology	Role
🧠 Google ADK	Multi-agent orchestration (ParallelAgent → SequentialAgent)
🎛️ Multimodal AI Agents	Parallel frame-level reasoning across pacing, visuals, and content
🐍 Python	Backend, frame extraction, and agent logic
🟢 JavaScript	Chrome extension — interception and live intervention
🌐 Django	Analysis & routing API
☁️ GCP	Hosting the agent service
💎 Gemini	Policy-synthesis agent
🤖 OpenAI	Pacing, visual-safety, and content-safety specialist agents
✍️ Prompt Engineering	Task-specific instructions and strict JSON contracts

---

🗺️ Roadmap & Future Scope

With a working application in place, MindSteady can be extended in several ways to maximize its utility and impact.

✅ Delivered

• [x] 🖥️ Chrome Extension with real-time video interception
• [x] ⚡ Live playback throttling & CSS dampening
• [x] 🤖 Google ADK multi-agent pipeline
• [x] 🌐 Django analysis & routing API
• [x] 🎨 Premium brand identity

🔭 Next

• [ ] 🎛️ Customizable Moderation — parents personalize filtering per child. If a child has a phobia (e.g. dogs), the parent specifies it and the AI flags or blocks videos containing that trigger to prevent unnecessary stress.
• [ ] 📱 Mobile Integration — extend beyond the browser into mobile environments using a "display over other apps" service, monitoring video across popular mobile apps with the same real-time protections.
• [ ] 📊 Real-time frame-level pacing telemetry
• [ ] 🔇 Audio spike detection & stabilization
• [ ] 👨‍👩‍👧 Parent dashboard with aggregate (non-identifying) insights
• [ ] 🌍 Multi-language safety policies
• [ ] 🧩 SDK for third-party platforms

By iteratively building on this foundation, MindSteady can grow into a comprehensive AI-powered safety layer for children's digital media consumption — bridging the regulatory gap with scalable, customizable protection.

---

🏆 Why MindSteady Wins

🎯 Judging Criteria — Mapped

Judging Criterion	Weight	How MindSteady Delivers
⚙️ Technical Execution	25%	A working monorepo: real DOM interception, live playback control, a Django API, and a functioning ADK multi-agent pipeline with caching and graceful degradation.
🤖 AI Relevance	20%	A genuine multi-agent architecture — parallel specialist agents plus policy synthesis — not a single wrapped API call.
🎯 Usefulness	15%	Solves the concrete "passive filter" gap: intervention, not just flagging.
🌍 Impact	15%	Research-grounded and scalable across every site with HTML5 video.
✨ Creativity	15%	The runtime-intervention model plus a closed-loop feedback architecture is genuinely differentiated.
🎤 Presentation	10%	Clear problem framing, architecture diagrams, and a live demo of real interventions.

🥇 The Differentiators

#	Differentiator	Why It Matters to Judges
1	Intervention, not blocking	Addresses the exact gap the brief calls out
2	True multi-agent design	Earns the AI-relevance weight honestly
3	Privacy-first	Matches the brief's "no excessive surveillance" stance
4	Research-grounded	Impact is defensible, not speculative
5	Working end-to-end MVP	Technical execution is demonstrable, not aspirational
6	Closes a documented regulatory gap	No existing tool moderates the visual characteristics research proves harmful

---

📅 Hackathon Timeline

Event: Child Safety in AI — hosted by STEMSisterz on Devpost Theme: Create apps that address the sociotechnical challenges of protecting children in the age of AI. Format: Online · Public · Students only · Beginner friendly

---

👤 Author

Muhammad Mujtaba Built for the Child Safety in AI hackathon — because a "safe" video should also be a gentle one.

---

🛡️👶 MindSteady

Protecting developing minds — one video at a time.

Built with ❤️ for the Child Safety in AI Hackathon

Built With

• ai
• child-safety
• chrome
• django
• gemini
• google-adk
• gpt-4.1
• javascript
• litellm
• manifest-v3
• multi-agent-systems
• privacy
• python
• real-time
• video