DoubleTake

A short-video app (think TikTok) that trains your eye to spot AI videos, powered by Nemotron agents

Comment 1

Inspiration

The inspiration for this family actually came from concern that AI content could maliciously misinform people I care about. AI has gotten exceptional at producing content at scale with very little effort. In the past, AI-generated content exhibited many artifacts that distinguished it from human-produced media. However, recent technological advancements have made the tell-tale signs of AI much harder to distinguish to the untrained eye. Fake videos of celebrities, politicians, car crashes, cats, and even advertisements are slowly corrupting social media algorithms and making their way onto our feeds. Although Gen Z might be able to distinguish current differences since we're growing up with this technology at our fingertips, it's easy to see that older generations are less able to detect AI-generated signs and are thus more susceptible to AI-generated scams and misinformation, which I witness firsthand with my parents being unable to tell what's AI and what's not. Now more than ever, it is imperative that we educate people on how to tell when videos are produced with AI to help prevent the spread of misinformation online.

What it does

  • This problem is why I built Double Take. Double Take is an AI detection short-form content web platform that gamifies the process of detecting common signs of AI-generated footage. Created with a familiar, fluid, reels-style user interface, Double Take provides a familiar experience in scrolling through videos on your favorite social media platforms; only it educates you on how to detect AI in the process. Users scroll through potentially AI-generated reels and hold or double-tap any clip to open Double Take: a spotlight that forwards/reverses to the right moment and eases onto the parts of the video that give away that it was produced by AI, offering a line-by-line explanation of why.
  • We also offer Video AI Detection, in which the DoubleTake Investigator, a Nemotron agent, plans which checks to run, calls tools, reacts when a step fails, and returns a verdict: Likely AI or Likely Real. With a "How we checked" timeline documenting every step it took.
  • A Skills view that shows your score, a range of graphs modeling different key AI notifiers, which are updated through every single reel you watch. With the ability to practice on reel sets that are personally chosen based on the notifiers you might be most susceptible to.
  • And full Account functionality complete with a guest mode, username sign-in, likes, comments, reposts, and your own personal profile.

How we built it

  • App: Expo SDK 58 / React Native + web, Expo Router, Reanimated. Every state change moves continuously (no snaps) and holds 60 fps on an iPhone in Expo Go.
  • Feed: 235 vetted real and AI clips with timestamped, boxed "tells", served from a Cloudflare Worker with signed media URLs.
  • Investigator (NVIDIA):
    • Nemotron (Bolt) plans and picks the next action from what it has seen so far, then selects the final evidence.
    • Nemotron Super VL inspects frames, Nano Omni describes the audio.
    • Cosmos-Reason2-8B, LoRA-fine-tuned locally on an RTX 4080 over 2,801 videos (about 45 minutes, $0), proposes where the tells are.
    • ffmpeg, C2PA (Content Credentials), and source checks.
    • Every step is recorded as a sanitized public trace (tool, why it ran, what it found, timing), which the app renders as "How we checked".
  • Backend: a private job runner bound to each upload's exact bytes, with quotas, caching, and recovery. Supabase handles accounts, cloud config, and prepared demo checks.
  • How we worked: one orchestrating Claude Code session ran many parallel agent threads (Claude and OpenAI Codex), each in its own git worktree, with an automated "hill climb" on the detection pipeline.
  • I was most proud of the system I devised to make claude code come up with multiple different design concepts, which I used to pick the best ones ## Challenges we ran into
    • AI Video Detection is Difficult (the highest accuracy attained from the best commercial detector got about 78% accuracy. So AI detection is proven to be challenging to make perfect.
  • Gathering quality training data and finding that fine line between mass-produced "slop" content and genuinely difficult-to-detect clips.
  • Hero Page was very demanding in trying to optimize for performance but retain visual appeal.
  • Remote computer shut down halfway through ## Accomplishments that we're proud of
    • I have an end project!
    • First SOLO hackathon
    • Working APP
    • Very proud of my UI

What we learned

  • Agents need tight budgets and visible traces to be trustworthy. Measuring honestly beats polishing a number. And motion design matters as much as model quality for whether people actually learn.

What's next for Project1

  • Better detectors for inserted subjects and synthetic voices
  • A "what to do next" flow for suspected scams
  • A share-sheet / browser extension to check a clip from any app
  • A public HTTPS analysis service.
  • Learning from players' answers to rank which clips teach best.
    • Incorporate into major social media apps (TikTok, Instagram)

Built With

Share this project:

Updates

Submission history