MindSafe

MindSafe is a full-stack system and Chrome extension that analyzes YouTube and YouTube Kids videos to estimate how healthy they are for a child’s mental and emotional development. It turns a complex AI pipeline into a simple 1–10 score and rich breakdown that parents can see directly on YouTube.

---

Why MindSafe Exists

Online video is a huge part of childhood today. It can be:

• Deeply positive: calm pacing, clear stories, kind language, strong social–emotional modeling.
• Quietly harmful: hyper-fast cuts, constant shouting, manipulative hooks, “brainrot” humor, or themes that are confusing or emotionally intense for young kids.

Most platforms focus on content category (kid vs. not kid) but not on the developmental quality of what a child is watching. A video can be “kid-friendly” and still be:

• Overstimulating for a 4‑year‑old.
• Emotionally confusing or scary.
• Teaching social dynamics that are harsh, sarcastic, or unkind.

MindSafe tries to fill that gap. It doesn’t replace parenting or clinical judgment, but it gives families a fast, evidence‑inspired signal about how a video might impact a young mind.

---

What MindSafe Does

• Analyzes YouTube / YouTube Kids videos given a URL and a child’s age.
• Downloads and processes the video (audio + visuals) locally for privacy.
• Transcribes the audio with a local Whisper model (no cloud upload of raw audio).
• Computes a set of metrics, including:
  • Overall Developmental Score (0–100) – higher is better.
  • Brainrot Index (0–100) – higher means more risk of negative impact (overstimulation, low-quality engagement).
  • Dimension scores:
    • Pacing (calm vs. chaotic)
    • Story / Narrative coherence
    • Language (kind vs. harsh/explicit)
    • Social‑Emotional Learning (SEL)
    • Fantasy balance (imagination vs. confusion)
    • Interactivity / call‑to‑action style
• Maps everything into a 1–10 MindSafe rating with an interpretation label like “Highly suitable”, “Moderately suitable”, or “Not recommended”.
• Displays the score directly in YouTube via a Chrome extension sidebar card and popup.

The goal is speed + decent accuracy: fast enough to use in real time, but grounded in real signals (transcripts, pacing metrics, heuristics, and optional LLM semantics) instead of random guesses.

---

High-Level Architecture

MindSafe consists of several coordinated parts:

• Chrome extension (chrome_extension/)

  • Injects a MindSafe card into YouTube / YouTube Kids watch pages.
  • Sends the current video URL to the local API.
  • Polls for the latest evaluation and renders:
    • Big 1–10 score badge (color‑coded green/orange/red).
    • Label (e.g., “Highly suitable”).
    • Expanded breakdown with all dimension scores.

• AI evaluation API (ai-agents/)

  • Flask API on http://localhost:5001 with endpoints like GET /evaluate?url=...&age=....
  • Downloads the video (via yt-dlp), extracts audio/video tracks, and runs the evaluation pipeline.
  • Uses local Whisper for transcription and a combination of:
    • Heuristic metrics (keyword patterns, pacing, structure).
    • Embeddings / semantic metrics (optionally via OpenRouter + Gemini models).
  • Returns a structured JSON object with all scores and metrics.

• Frontend + backend web app (frontend/ + backend/)

  • Minimal Flask app that serves a landing page and basic UI.
  • Useful for demos outside the browser extension context.

---

Technical Stack

• Languages & Frameworks

  • Python (Flask API, AI pipeline)
  • JavaScript (Chrome extension, frontend behavior)
  • HTML/CSS (popup UI, sidebar styling)

• AI / Data Processing

  • Whisper (local) for speech‑to‑text transcription (no raw audio sent to the cloud).
  • yt-dlp for reliable YouTube / YouTube Kids video download.
  • FFmpeg for audio and video track extraction.
  • Heuristic and embedding‑based metrics for semantics and narrative coherence.
  • OpenRouter + Gemini (optional, via evaluation/llm_client.py) for deeper semantic labeling.

• Infrastructure / Integrations

  • Snowflake (optional) for logging evaluation results at scale (video URL, scores, full JSON result) so you can build dashboards or run research on how kids’ content has changed.
  • Vultr cloud (optional) as a remote worker target:
    • You can run a copy of the ai-agents API on a powerful Vultr instance.
    • A local API can forward evaluations to the Vultr worker when configured, speeding up scoring for heavier models.

---

End-to-End Workflow

1. User watches a video

• The user opens a YouTube or YouTube Kids watch page.
• The MindSafe content script detects that this is a /watch page and:
  • Injects the MindSafe card into the right column.
  • Sends a NEW_VIDEO message with the video URL to the background service worker.

2. Background service worker calls the API

• The background script:
  • Stores a pending lastScore (status pending) in memory and in chrome.storage.local so the UI can immediately show “Analyzing…”.
  • Calls http://localhost:5001/evaluate?url=...&age=... with the configured child age.

3. AI pipeline runs (locally or on Vultr)

Inside ai-agents/api.py and evaluation/:

• Download & extract

  • yt-dlp downloads the video.
  • ffmpeg extracts:
    • video_with_audio.mp4
    • audio_only.m4a
    • video_no_audio.mp4 (muted video)

• Transcription (Whisper)

  • The audio is chunked and transcribed locally using Whisper.
  • Chunks are combined into a full transcript for downstream analysis.

• Metrics & scoring

  • Audio pacing and rhythm features.
  • Simple shot/segment structure (time‑based segmentation).
  • Heuristic metrics on the transcript (e.g., intensity, language style, SEL cues).
  • Optional LLM‑based semantics via OpenRouter if enabled.
  • All metrics are combined into:
    • dev_score (0–100).
    • brainrot_index (0–100).
    • dimension_scores for Pacing, Story, Language, SEL, Fantasy, Interactivity.
    • A 1–10 MindSafe rating + human‑readable label.

• Optional: Cloud acceleration & logging

  • If configured, a remote Vultr worker can handle the heavy evaluation instead of your local machine.
  • Snowflake can asynchronously receive a copy of each result for analytics and research.

4. Results flow back to the extension

• The API returns JSON to the background service worker.

• The background script:

  • Enriches the result with label + reasons.
  • Stores it as the new lastScore (in memory + chrome.storage.local).
  • Broadcasts a SCORE_UPDATED message to all YouTube tabs.

• The content script:

  • Receives SCORE_UPDATED and re‑renders the MindSafe card.
  • Shows a big X / 10 score badge with colors:
    • Green: generally very supportive and age‑appropriate.
    • Orange: mixed, good with some caution or supervision.
    • Red: likely overstimulating, confusing, or emotionally intense for young kids.
  • On click, it expands to show all dimension scores and context.

• The popup:

  • Reads lastScore from storage and shows a compact dashboard for the last analyzed video.

---

Why This Matters for Children’s Mental Health

Children’s brains are still wiring up their attention, emotion regulation, and understanding of relationships. Video content can either:

• Support that development: gentle pacing, clear cause‑and‑effect, kind language, healthy problem‑solving.
• Undermine it: constant dopamine spikes, aggressive or mocking social dynamics, confusing fantasy vs. reality, or “always on” stimulation with no downtime.

MindSafe is built around that reality:

• It does not just ask “Is this kid content?” but “Is this likely to be mentally healthy for this specific age?”
• It treats brainrot not as a joke but as a rough index of:
  • Overstimulation and frenetic pacing.
  • Low‑quality engagement vs. learning or reflection.
  • Repeated patterns that may encourage addictive viewing rather than balanced habits.

This tool is not a diagnosis or a replacement for parenting, but it can:

• Help parents quickly triage what kids are watching.
• Start better conversations about “why this video feels good or not‑so‑good”.
• Provide researchers and clinicians (via logs, if enabled) with a starting point for studying the impact of modern kids’ media.

---

Getting Started (High-Level)

At a high level you will:

1. Set up the AI API (ai-agents/)

   • Create and activate a Python virtual environment.
   • Install dependencies from ai-agents/requirements.txt.
   • Configure .env (OpenRouter key if using LLM, optional Snowflake / Vultr envs).
   • Run python api.py to start the Flask server on http://localhost:5001.

2. Load the Chrome extension (chrome_extension/)

   • Go to chrome://extensions → enable Developer mode.
   • Click Load unpacked and select the chrome_extension folder.

3. Open YouTube / YouTube Kids and watch a video

   • The MindSafe card appears in the right column.
   • After processing, you’ll see the 1–10 score and can expand the details.

For more technical detail and API usage, see the documentation inside ai-agents/ (e.g., API_README.md, RUN_API.md, and MINDSAFE_SCORES.md).

---

Future Potential

MindSafe is intentionally modular and has room to grow:

• Smarter models: swap in stronger LLMs or domain‑specific models as they emerge.
• Personalized profiles: tune thresholds for specific ages, sensitivities, or neurodivergent kids.
• Richer dashboards: leverage Snowflake logs to build visuals showing trends over time.
• Cross‑platform support: extend the extension and pipeline beyond YouTube / YouTube Kids.

Most importantly, it is a step toward treating kids’ mental health as first‑class in product design—not an afterthought.