RealSight

Passive AI-video detection, right on the player, as you browse.

Comment

RealSight

Passive AI-video detection, right on the player, as you browse.

Inspiration

AI-generated video is flooding YouTube and Instagram, and every existing detector makes the viewer do the work: right-click a thumbnail, paste a URL into a web tool, hope someone crowd-flagged it. Detection that requires effort protects nobody at scale. We asked: what if checking were as passive as watching?

What it does

RealSight is a Chrome extension that automatically analyzes every video you watch on YouTube (including Shorts) and Instagram Reels. Within ~5 seconds of playback, a badge appears on the player: 🤖 AI Generated (87%), ✅ Likely Real (95%), or ❓ Uncertain — with the model's specific visual reasoning (e.g., "over-smooth, plastic-looking skin", "repetitive, unnatural patterns on clothing") one click away in the popup. No clicks, no copy-paste, no manual checks. The toolbar icon mirrors state in real time: gray off-page, pulsing blue while analyzing, verdict-colored when done.

How we built it (Technological Implementation)

  • Google ADK 2.x agent + Gemini (gemini-2.5-flash) hosted on Google Cloud Run. The agent receives 3–4 JPEG frames extracted client-side from the playing <video> via canvas (content scripts run same-origin, so the canvas never taints — this also works on Instagram's MSE blob streams).
  • Official MongoDB MCP Server (mongodb-mcp-server) as the agent's tool — the partner integration. The agent's first action on every request is a find against the realsight.detections collection in MongoDB Atlas, keyed on a normalized video URL (watch pages, Shorts, youtu.be links, and Reels permalinks all canonicalize to one key). A cache hit returns the verdict in under a second without invoking Gemini at all — cheaper, faster, and verdicts are consistent for every user who watches the same video. On a miss, Gemini performs forensic frame analysis (texture smearing, anatomy errors, lighting/shadow physics, temporal inconsistency between frames) and the agent writes the verdict back through the MCP server's insert-many tool.
  • Engineering details judges can check in the repo: strict-JSON agent output with validation and clamping; Pydantic request limits (≤5 frames, ≤1MB each); 503-retry with exponential backoff to survive Gemini load spikes; the Docker image pre-bakes Node + the MCP server so Cloud Run cold starts never hit the npm registry; the extension is vanilla MV3 JavaScript with zero dependencies and zero build step.

Design / UX

The product goal was zero-interaction detection: the verdict appears where your eyes already are (the player), in glanceable color (red/green/gray), with confidence stated as a percentage and reasons in plain language. Failures degrade gracefully — an unreachable backend yields a dismissable "analysis failed" badge, never a broken page. The animated toolbar icon communicates state even when the badge is dismissed.

Potential Impact

Deepfake detection tools today serve journalists and researchers — people who already suspect something. RealSight targets the other 99%: passive viewers who never think to check. Ambient, automatic labeling is how media literacy scales. The MongoDB cache makes this economically viable: each video is analyzed once globally, then served from Atlas for every subsequent viewer.

Honest limitations & what's next

Verdicts are Gemini-based visual heuristics, not watermark forensics, and we present them as confidence levels, not facts. The detector sits behind an interface designed for Google's SynthID Content Detection API (currently partner-preview) to slot in when publicly available. Also next: frame-hash dedup to catch re-uploads under new URLs, sensitivity settings, and a Chrome Web Store listing.

Hackathon compliance checklist (for judges)

  • ✅ Agent built with Google ADK + Gemini, hosted on Google Cloud Run
  • ✅ Partner MCP server: official MongoDB MCP Server wired into the agent as its data tool → MongoDB partner bucket
  • ✅ Public repo, MIT license · hosted URL · ~3 min demo video

Demo video script (~3 min)

  1. The problem (20s) — AI video is flooding feeds; existing detectors need manual checks.
  2. Live detection (60s) — browse YouTube: AI-generated video gets 🤖 badge, real video gets ✅, popup shows Gemini's reasons.
  3. The MongoDB moment (30s) — revisit the same video: instant cached verdict via MongoDB MCP Server find against Atlas. Linger here.
  4. Architecture (45s) — Chrome MV3 extension → Cloud Run → Google ADK agent → Gemini (frame forensics) + MongoDB MCP Server (cache) → Atlas.
  5. Impact (15s) — passive media literacy for everyone; SynthID Content Detection API as the upgrade path.

Built With

Share this project:

Updates