Home Page

PlayProof

Inspiration

We are living under what many call the “Dead Internet,” where the internet mainly consists of bot activity and automatically generated content. A dead internet isolates and manipulates individuals, causing many to experience financial and emotional burdens.

30% of all online reviews are from bots, resulting in $787B in wasted consumer spending.
49.6% of all social media accounts are bots, wasting over $41.4B in ads from individual entrepreneurs each year.
There is a 333% increase in bots on romance apps in 2025, resulting in $1.14B lost to romance scams.

Why is this happening? Models are getting smarter.

YOLOv8 could beat Google’s ReCaptcha v2 100% of the time. Almost every LLM Agent gets 60-90% of human verification correct. This level of accuracy will only keep growing as LLM Agents get better, to the point where status quo human verification will be pointless.

Current human verification needs to change, and PlayProof is here to define that change.

What it does

Playproof is a universal, plug-and-play SDK that enables human verification through AI-generated games. Organizations can embed branded verification experiences that utilize:

Fast-paced games: Simple games that require human-like reaction time and movement
ML-powered detection: Integrates with Woodwide for anomaly detection using granular movement features (velocity, acceleration, jerk, path efficiency, jitter) to distinguish humans from bots
Smart observability: Monitors gameplay events and regenerates new games if suspicious patterns are detected
Seamless branding: Customize colors, sprites, backgrounds, and visual elements to match brand identity and offer seamless experiences to users.
3D game engine: Built with Three.js for engaging 3D experiences (with 2D orthographic options)

After 300 tests with the latest AI models, it successfully defends against 99.9% of attacks, with the longest response time in minigames at 10 seconds.

How we built it

The architecture consists of:

SDK (TypeScript + Three.js): Client SDK with a Three.js game engine, real-time telemetry tracking, and LiveKit integration for observability. Games run in the browser and stream behavioral data in real time.
Backend (Next.js): API routes with Woodwide ML integration for scoring and anomaly detection. Handles feature extraction, model inference, and decision logic.
Database (Convex): Real-time backend for deployments, branding configurations, and verification results. Provides instant updates and low-latency queries.
Dashboard (React + Next.js): Admin interface for managing deployments, viewing verification attempts, and testing games with live telemetry visualization.

Challenges we ran into

User Experience: It was incredibly difficult to create an experience that users wouldn't find annoying or dislike. We spent hours debating on how the SDK should operate, how long it should last, how games should be generated, and more.

Game Generation: The game generation itself was extremely challenging. The system required constantly changing games generated by LLMs, but this posed a challenge because LLMs could introduce errors without proper validation. We overcame this challenge by gradually building the game scene through chained tool calls rather than a single prompt. Each tool call validated what was already in the scene, with no errors.

Real-time ML Inference: The WoodWide AI API had limited credits, which caused trouble when a user had only 50 credits, and each inference cost 1 credit. To handle this, we created a mixture of heuristics based on consistency and smoothness, alongside our trained ML model in WoodWide, to efficiently handle credit transfers between team members.

Accomplishments that we're proud of

Extremely accurate: The bot detection works extremely well, and our system defended against almost every test case. We believe we've figured out the next wave of human verification!

Universal SDK: The SDK is extremely simple and works anywhere. This makes it easy enough to use for almost every platform.

End-to-End System: The application is fully usable from end to end. Someone can create a deployment, brand their organization, and utilize the human verification already!

What we learned

Behavioral patterns matter: Mouse movement features (jerk, path efficiency) are more reliable indicators than simple click accuracy

Hybrid ML approaches work: Combining real-time heuristics with batch ML inference provides both speed and accuracy.

User experience is key: Game interfaces are more engaging than traditional CAPTCHAs, improving completion rates.

What's next for Playproof

More game types: Adding rhythm games, puzzle games, and other engaging verification experiences

Mobile and gaming support: Extending to touch-based games for mobile verification, as well as video games to combat the AI bot epidemic in gaming!

Advanced ML models: Training custom models on our collected data for improved accuracy

Enhanced branding system: Supporting custom logos, background images, game sprites, and branded assets so organizations can fully customize the visual experience beyond colors and typography