🛡️ The world's first 100% client-side, science-grounded engine that quantifies sensory overload in children's media.
Analyze any video in your browser and get a science-based age recommendation — without a single byte ever leaving your device.
-FFD21E?style=flat-square){kind=icon}
⬆️ Drag & drop a video → get an objective, science-backed age verdict. Screenshots from the live app.
This project was designed and built by Daniyal Khan exclusively for the Devpost Child Safety in AI hackathon. It is a purpose-built submission, not a repurposed project — every feature below was engineered within the hackathon window to answer the challenge brief.
| 🎯 Hackathon | Child Safety in AI |
| 🏫 Hosted by | STEMSisterz |
| 🗓️ Build window | October 13, 2026 @ 6:00 AM GMT+5 → October 15, 2026 @ 6:00 AM GMT+5 |
| ⏱️ Deadline | Oct 15, 2026 @ 6:00 AM GMT+5 |
| 👥 Participants | 48 |
| 🏅 Prizes | 🥇 $250 Winner · 🥈 $150 Runner-up (Amazon gift cards) |
| 🧭 Category | Machine Learning / AI · Beginner Friendly · Low/No Code |
| 👩⚖️ Judge | Chloe Lin |
The brief asked for solutions to the sociotechnical challenges of protecting children in an age of AI. Kiddometer tackles three of the listed areas head-on:
| Challenge area from the brief | How Kiddometer delivers |
|---|---|
| ✅ AI-powered content safety — detect or filter harmful / age-inappropriate content | Automated computer-vision + audio-signal analysis flags sensory-toxic content that human ratings miss |
| ✅ Parental and educator tools — support safety without excessive surveillance | Fully client-side; parents get guidance from their own file, no accounts, no tracking, no data collection |
| ✅ Trustworthy AI systems — evaluate, audit, or improve the safety of AI products used by children | An auditable, open-source, peer-reviewed-methodology scoring engine anyone can inspect and re-run |
Parents are told a video is "safe" because there's no bad language and no violence — but the things that actually reshape a toddler's developing brain are invisible to every traditional rating:
- 🔪 Rapid scene cuts every 1–3 seconds
- 🔊 Sudden loudness spikes ("jumpscares" and sound stings)
- ⚡ Flashing and strobing that breaches seizure-safety limits
- 🌈 Extreme saturation and edge density that overloads young visual systems
Age ratings like TV-Y and PG judge plot and profanity. They are blind to sensory overload, hyper-pacing, and flashing. Peer-reviewed research links fast-paced editing to measurable executive-function deficits in preschoolers — yet no consumer tool quantifies it.
Kiddometer fixes that.
Drop in a video (MP4 / WebM / MOV), paste a direct URL, or use the optional YouTube tab. In roughly ¼ of the video's runtime, Kiddometer decodes the audio-visual stream entirely in your browser and returns an objective, science-based age recommendation.
The video never leaves your device. Decoding, frame inspection, and signal analysis execute 100% locally — eliminating COPPA and GDPR-K compliance risk by construction.
Every score is traceable to an established, peer-reviewed standard. No subjective human opinion — quantitative signal telemetry.
| 📊 Metric | 🧠 What it detects | 📚 Scientific basis |
|---|---|---|
| Visual cuts/min | Shot-change frequency via luminance-histogram correlation | Lillard & Peterson (2011); Hinten et al. (2025) executive-function thresholds |
| Flashes/s | Per-zone luminance oscillations | WCAG 2.3.1 photosensitive-seizure limit (3 flashes/s) |
| Loudness peaks | Short-term loudness jumps vs. programme average | ITU-R BT.1770 K-weighting; EBU R128; WHO safe listening (<75 dB) |
| Sound event density | Salient audio onsets per minute | Christakis et al., PNAS (2018) |
| On-screen motion | Inter-frame luminance delta within shots | Calibrated heuristic |
| Visual complexity | Sobel edge density per frame | Calibrated heuristic |
- The video is split into 60-second segments.
- Each segment gets the age band its strictest metric demands.
- The global rating is the 90th percentile across segments — sustained intensity decides, so a single outlier peak can't inflate or deflate the verdict (it's surfaced separately as an intense moment).
- Any breach of the 3 flashes/s limit always raises an explicit
⚠️ photosensitivity warning.
Age bands: 0–3 · 4–6 · 7–12 · 13+
Kiddometer is a zero-build, zero-dependency vanilla-JS application. All heavy signal processing runs in-browser using native platform APIs — no WebAssembly blobs, no frameworks.
flowchart TB
subgraph Client["🌐 Browser — 100% client-side (nothing uploaded)"]
direction TB
UI["🖥️ index.html + styles.css<br/><i>Design tokens · Basic / Expert modes</i>"]
MAIN["🎛️ main.js<br/><i>UI orchestration & rendering</i>"]
INS["🧠 insights.js<br/><i>Data-driven verdict, advice & tips</i>"]
SCORE["📐 scoring.js<br/><i>Age thresholds · 90th-percentile aggregation</i>"]
subgraph Engines["🔬 Analysis engines"]
direction LR
VID["🎞️ video-analyzer.js<br/><i>Canvas frame extraction<br/>RGB→Luminance · histogram correlation<br/>Sobel edges · WCAG flash detection</i>"]
AUD["🔊 audio-analyzer.js<br/><i>Web Audio / OfflineAudioContext<br/>BT.1770 K-weighting<br/>short-term loudness · onset detection</i>"]
end
REPORT["📄 Self-contained HTML report<br/><i>timelines · donuts · sparklines</i>"]
end
subgraph Optional["☁️ Optional backend (Hugging Face Space)"]
BE["🐍 FastAPI + yt-dlp<br/><i>stateless byte relay ≤480p<br/>stores nothing</i>"]
end
SRC{{"📥 Video source"}}
SRC -->|local file / direct URL| MAIN
SRC -->|YouTube / Vimeo / TikTok| BE
BE -.->|streams bytes, no storage| MAIN
MAIN --> VID
MAIN --> AUD
VID --> SCORE
AUD --> SCORE
SCORE --> INS
INS --> REPORT
UI <--> MAIN
| Module | Responsibility | Key techniques |
|---|---|---|
docs/js/main.js |
UI orchestration, source handling, rendering | requestVideoFrameCallback, drag & drop, view state |
docs/js/video-analyzer.js |
Cuts, motion, complexity, flashes | HTML5 Canvas, RGB→luminance, 32-bin histogram correlation, Sobel edge filter, 3×3 flash zones |
docs/js/audio-analyzer.js |
Loudness peaks, sound-event density | Web Audio API, OfflineAudioContext, BT.1770 K-weighted biquads (highpass + high-shelf), 400 ms short-term windows, energy-flux onsets |
docs/js/scoring.js |
Age thresholds & aggregation | 60 s windows, strictest-metric classification, 90th-percentile rollup, photosensitivity escalation |
docs/js/insights.js |
Human-readable verdict & guidance | Generated from the actual measurements |
docs/js/config.js |
Optional backend URL | Single line — empty disables the YouTube tab |
| Feature | |
|---|---|
| 🗂️ | Three input sources — local file, direct video URL, or YouTube/Vimeo/TikTok via the optional backend |
| 🖱️ | Drop & scan — drag a video in; analysis takes ~¼ of its duration |
| 🧒 | Basic / Expert modes — a plain verdict for parents, full charts for the curious |
| 📈 | Rich diagnostics — per-minute timeline, intense moments, distribution donut, per-metric sparklines |
| 📝 | Downloadable report — self-contained HTML with the full per-segment table |
| 🗣️ | Data-driven copy — verdict summary, advice and tips are generated from the real measurements |
| 🔐 | Private by design — analysis is 100% client-side; the optional backend only relays bytes and stores nothing |
| ⚡ | Zero dependencies / no build step — clone and serve |
How Kiddometer maps to each weighted criterion:
| Criterion | Weight | Where Kiddometer scores |
|---|---|---|
| Technical Execution | 25% | In-browser Canvas frame decoding, Web Audio API signal processing, and a zero-build vanilla-JS architecture — complex DSP (BT.1770 K-weighting, Sobel edge detection) implemented natively with no heavy WASM libraries |
| AI Relevance | 20% | Automated computer-vision heuristics (luminance-histogram correlation, Sobel edge density) + automated audio feature extraction (short-term loudness, onset detection) replace manual review |
| Usefulness | 15% | Gives parents instant, actionable age guidance before a child watches — no account, no upload, no waiting |
| Impact | 15% | Zero-latency, 100% private, COPPA-compliant by design; scales to any browser on earth and into extensions, smart-TV firmware and school dashboards |
| Creativity | 15% | Shifts content rating from subjective human review and text tags to quantitative sensory-signal telemetry |
| Presentation | 10% | Clean Basic/Expert UI, self-contained report, and a concise demo walkthrough |
Open the app, drop a video, and wait roughly ¼ of the video's duration. The video never leaves your device — decoding, audio analysis, and scoring all run in the browser.
git clone https://github.com/digiflowhelp-stack/Kiddometer
cd Kiddometer/docs
python -m http.server 8000
# → http://localhost:8000No build step. No dependencies. Just a browser (Chromium-based recommended for full requestVideoFrameCallback support).
| Layer | Technology |
|---|---|
| Frontend | Pure HTML5 · CSS design tokens · Vanilla JavaScript (ES modules) |
| Video analysis | HTML5 Canvas API · requestVideoFrameCallback · luminance histograms · Sobel edges |
| Audio analysis | Web Audio API · OfflineAudioContext · ITU-R BT.1770 K-weighting · EBU R128-style loudness |
| Scoring | Custom 90th-percentile windowed aggregation engine |
| Optional backend | Python · FastAPI · yt-dlp · Docker (Hugging Face Spaces) |
| Legacy reference | Streamlit (app.py, core.py) |
Kiddometer
├── docs/ # 🌐 Static web app (GitHub Pages)
│ ├── index.html # Home, analysis & methodology views
│ ├── assets/ # Logo (SVG)
│ ├── css/styles.css # Design tokens (from design/desgin.pen)
│ └── js/
│ ├── main.js # 🎛️ UI orchestration and rendering
│ ├── insights.js # 🧠 Data-driven copy (verdict, advice, tips)
│ ├── video-analyzer.js # 🎞️ Cuts, motion, complexity, flashes
│ ├── audio-analyzer.js # 🔊 BT.1770 loudness, onset detection
│ ├── scoring.js # 📐 Age thresholds and aggregation rules
│ └── config.js # ⚙️ Optional backend URL
├── backend/ # ☁️ Optional download helper (HF Space)
├── design/ # 🎨 Pencil (.pen) design source
├── app.py / core.py # 🕰️ Legacy Streamlit version
└── .github/screenshots/ # 🖼️ App screenshots
Streaming platforms don't allow fetching video files cross-origin, so the YouTube tab uses a small FastAPI + yt-dlp helper that streams the file to the browser (≤480p, nothing stored). It runs free on a Hugging Face Space and includes anti-bot layers for datacenter IPs: PO Token provider, Chrome TLS impersonation, a client fallback chain, and 12-hourly yt-dlp self-updates. Optional cookies act as the last layer.
Deploy guide: backend/DEPLOY_HF.md. Point docs/js/config.js at your Space and you're done. Without a backend, everything else still works.
#ChildSafety #ResponsibleAI #AIAct #COPPA #GDPR-K #PrivacyByDesign #ComputerVision #WebAudio #SignalProcessing #EdTech #DigitalWellbeing #SensoryOverload #WCAG #ITU-R-BT.1770 #OpenSource #VanillaJS #ClientSideAI #STEMSisterz #Devpost #Hackathon
See CONTRIBUTING.md. Issues and PRs welcome — especially calibration data for the heuristic thresholds.
MIT — Kiddometer is an orientation tool; it does not replace the judgement of parents or professionals.
Built with ❤️ and a lot of ☕ by Daniyal Khan for the Child Safety in AI hackathon (Oct 13 – Oct 15, 2026).
🛡️ 100% local · 🚫 0 bytes uploaded · 📖 peer-reviewed methodology · 🌍 built for every child's browser