truthful

Truthful, an end-to-end security and fraud detection platform with a real-time motion-physics AI detector that’s plug-and-play for creators, users, and businesses across any app, or browser.

Comment

Inspiration

In an era of digital evolution it can feel like technology is advancing faster than what the average user can keep up with as over 85% of people around the world receive digital scam attempts. Its clear that the world needs a security platform thats safe, realiable and trustworthy while also catering towards new technologies such as the evolution of AI videos.

Moderators, educators small creators and businesses need a quick, plug-in safety net—not direct experience in AI and machine learning. This is where truthful comes in.

What it does

Easy to use AI detection, just upload a video or paste a youtube link to get a confidence score and rating provided by our model, you can also track your history within the application to see previously analyzed videos as well secure account creation storage.

  • Skeleton-Motion AI Detector – extracts joint trajectories, measures physics-consistency, flags deepfakes in <300 ms.
  • REST API + Webhooks – one POST and you get "is_ai": true|false with a confidence score.
  • Render disk deploy – weights load instantly; zero cold-start hit.
  • Frontend Demo – upload → live gauge flips from Real to AI.

How I built it

Layer Stack
Model ConvNeXt-Tiny backbone → 1-D Temporal CNN → skeleton-distance fusion (PyTorch).
Training 12 K clips, balanced sampler, focal loss → AUC 0.83 after 4 epochs.
Backend FastAPI, uvicorn, Docker → Render; model on /var/models disk.
Dev Ops GitHub → Render CI, Git-LFS for 335 MB pt; Netlify for React UI.
QA cURL smoke tests, TensorBoard + skeleton overlay heat-map.

Challenges I ran into

  • Data-Loader Freezes – PyTorch workers dead-locked at val batch #188; fixed with val_num_workers=0.
  • Class Imbalance – 1 : 11 AI/Real; had to whip up weighted sampler & pos_weight or accuracy flat-lined.
  • Fusion Collapse – CNN logits swamped by skeleton path; rescued AUC by warm-starting fusion weight at 0.5.
  • Late-night Ops – 350 MB model wouldn’t fit git plain; moved to Render disk with custom SSH key under deadline.

Accomplishments that I am proud of

First accomplishment being this is my first hackathon let alone project in AI/ML in general. I've always wanted to break into the industry and truthful feels like the perfect opportunity to do so. Im proud of the UI/UX design layout as well as the concrete model features and ease of use.

  • 0 → 0.83 AUC in 30 hrs – legitimate detection curve, no smoke-and-mirrors.
  • Sub-second inference on a CPU-only dyno.
  • One-liner integration – drop SKELETON_MODEL_PATH env-var and you’re live.

What I learned

  • Motion physics beats pixel-level cues for many generative errors.
  • Disk-based weight mounting on Render slashes cold-starts vs baking into the image.
  • Always log both per-class metrics and fused metrics—otherwise you’ll debug the wrong thing at 3 a.m.
  • A tiny React bar graph + clear copy builds more trust than piles of raw JSON.

What's next for Truthful

  1. Multi-modal stack – add audio & transcript alignment, watermark scan (preprocessing).
  2. Fraud detection – Add other measures of fraud detection that truthful can monitor in the back while clients continue on their day to day. This could include scam email detection, web fraud, and payment fraud.
  3. In app integration – add the ability to integrate truthful into business applications
  4. Chrome / Edge extension – on-device frame sampling → cloud verify.
  5. Creator dashboard – batch reports, dispute flow, SOC-2 logs.
  6. SDK & Pricing tiers – free 50 calls/day, $10 dev tier, enterprise SLA.
  7. Explainable overlay – show which joints break gravity so users see the glitch.

Truthful starts as a fast AI-video lie detector, but will grow into a full security toolkit anyone can put onto an app, site, or browser.

Built With

Share this project:

Updates