Privacy-Protective Video Playback Assessment for PWA/Android

DIGNARIA — Escudo Digital: On-device visual safety feasibility for PWA/Android

We need an experienced computer-vision and web/mobile engineer to assess and prototype a privacy-preserving, on-device visual safety layer for protected video playback.

Current state:
- Existing draft implementation only provides a static fail-closed policy and blocks uninspectable external playback.
- There is no production nudity classifier, no model weights, no Canvas/OffscreenCanvas pipeline, and no claim of device-wide protection.
- The contract must remain fail-closed: any invalid output, timeout, unavailable frame, CORS/DRM restriction, or runtime failure must return UNVERIFIED and keep video paused, muted and hidden.

First paid phase only — fixed-price feasibility spike:
1. Identify a technically viable local model or alternative training strategy whose code, weights and training-data provenance are legally usable and documented.
2. Do not use cloud moderation APIs. No image/frame/video may leave the device.
3. No persistence of frames, thumbnails, biometric data or raw URLs.
4. Build an isolated proof of concept for inspectable HTML5 video only, preferably using a Web Worker and WebAssembly/WebGPU where supported.
5. Provide reproducible dependency versions, licenses, SHA-256 checksums and a software/model bill of materials.
6. Provide network traces showing zero external traffic after local assets are installed.
7. Benchmark on at least one real mid/low-range Android device: startup time, inference latency, memory and thermal behaviour.
8. Test fail-closed behaviour for missing model, timeout, NaN/invalid predictions, sparse/duplicate/unknown classes, invalid cardinality, probability-sum errors, CORS/DRM/uninspectable content and worker/runtime failure.
9. Use only synthetic, institutional or clearly licensed non-explicit test material. Absolutely no minors and no real explicit material.
10. Deliver source code, test results, limitations and a short recommendation. Do not deploy or merge to production.

Important limitations:
- A PWA cannot control other apps, external browsers, private/incognito mode, VPN, DNS, downloaded files or casting.
- Third-party iframes and DRM content may be uninspectable and must remain blocked inside DIGNARIA.
- False positives in medicine, art, breastfeeding, sport and beach contexts must be measured and reported.
- Do not propose NSFWJS or another model unless the license of the actual weights and provenance/legal basis of the training data are evidenced.

Please include in your bid:
- Relevant experience with TensorFlow.js, ONNX Runtime Web, TFLite, Web Workers/OffscreenCanvas and Android WebView/PWA.
- The exact model or approach you recommend and documentary evidence for code, weights and dataset rights.
- Fixed price and delivery time for this feasibility spike.
- One example of a comparable on-device computer-vision project.
- Confirmation that you will not use cloud APIs or retain visual data.

We will shortlist only candidates who answer these points specifically. Generic SEO, WordPress, marketing or chatbot proposals will be rejected.