Orislop

Inspiration

Short-form platforms used to feel like places where people discovered genuinely interesting, original content. Over time, many feeds became overwhelmed by repetitive clips, AI-generated videos, recycled compilations, synthetic voices, engagement bait, and other low-effort content.

Most existing tools focus on blocking entire platforms or enforcing screen-time limits. We wanted to solve a different problem: let people keep using the platforms they enjoy while automatically filtering out the content they do not want to see.

That led us to build Orislop — a system designed to detect and skip low-value “slop” while preserving original and worthwhile content.

What it does

Orislop is a browser extension and AI-powered content-filtering system for short-form video platforms such as YouTube Shorts, TikTok, and Instagram Reels.

As a user scrolls, Orislop analyzes each video using signals such as:

Repetitive or low-effort visual patterns
AI-generated or synthetic content
Text-to-speech and artificial voices
Audio and video synchronization
Reposted or compilation-style content
Temporal patterns across video frames
Other indicators of low originality or engagement bait

The system assigns each video a classification and confidence score. Depending on the user’s selected strictness level, Orislop can automatically skip the video, mark it as questionable, or allow it to play normally.

Users can still scroll backward and view anything that was skipped, keeping them in control of their feed rather than permanently censoring content.

How we built it

We built Orislop as a combination of a Chrome extension, a desktop inference system, and a multimodal machine-learning pipeline.

The extension observes short-form video pages, detects when the active video changes, extracts the relevant media information, and sends it to the classifier. A look-ahead system can begin analyzing upcoming videos before the user reaches them, allowing unwanted content to be skipped with minimal delay.

For the machine-learning system, we developed multiple specialized detection components rather than relying on a single generic model. These include temporal models that analyze patterns across different time intervals, along with planned spatial, audio, and audio-video synchronization detectors.

Our temporal model uses several experts that analyze micro-, medium-, long-, and extra-long-term patterns within a video. Their outputs are fused and calibrated to produce a final confidence score.

The surrounding application was built using technologies including JavaScript, TypeScript, Electron, Chrome extension APIs, Python, PyTorch, CUDA, and local inference adapters. We also created configurable thresholds, per-signal explanations, recent-detection history, privacy controls, and fail-safe behavior when a video cannot be analyzed correctly.

Challenges we ran into

One of the largest challenges was defining “slop” in a way that could be consistently understood by both humans and machine-learning models. Low-quality content is not a single visual category. It can involve editing style, repetition, audio, originality, context, or combinations of several weak signals.

Dataset creation was another major challenge. There is no large, reliable public dataset labeled specifically for short-form slop. Human labeling is expensive and subjective, while using an older AI model to generate labels risks teaching the new model the same mistakes and biases.

We also had to balance accuracy with speed. A highly accurate model is not useful if the user has already watched several seconds of a video before it finishes analyzing it. This required us to experiment with lightweight local inference, heavier cloud-based analysis, caching, look-ahead scanning, and multiple model sizes.

Integrating with constantly changing social-media interfaces was also difficult. Each platform has different page structures, video-loading behavior, navigation systems, and restrictions. The extension therefore needed platform-specific adapters and defensive logic rather than a single universal implementation.

Finally, we had to avoid making the filter overly aggressive. False positives are especially damaging because they can cause original content to be skipped. We added adjustable strictness levels, confidence thresholds, explanations, and the ability to scroll back to give users greater control.

Accomplishments that we're proud of

We are proud that Orislop progressed beyond being only an idea and became a functional end-to-end prototype.

We built:

A working browser-extension interface
Automatic detection and skipping behavior
Support for multiple short-form platforms
A configurable strictness system
Per-signal classification explanations
Look-ahead scanning and pre-classification
A multimodal detection architecture
A temporal model composed of four specialized experts
Local inference adapters and privacy-focused controls
Safety mechanisms that prevent infinite skipping or navigation loops

During early real-world testing, improvements to model calibration and detection thresholds increased performance substantially compared with our earlier prototype. More importantly, we created an architecture that can continue improving as additional spatial, audio, synchronization, provenance, and user-feedback signals are added.

What we learned

We learned that content quality cannot be reliably determined from a single frame or signal. A convincing system needs to understand how visuals, audio, editing patterns, repetition, and context interact over time.

We also learned that the dataset may become more valuable than any individual model. Models can be replaced or improved, but a carefully labeled, diverse, and continuously updated dataset of real short-form content can become a long-term technical advantage.

Another major lesson was that user control is essential. Different people have different definitions of low-value content. Rather than forcing one universal standard, Orislop needs to learn from explicit user preferences while still providing sensible default behavior.

We also discovered that product engineering is just as important as model accuracy. Even a strong classifier can create a poor experience if inference is slow, videos are skipped incorrectly, or platform integrations break.

What's next for Orislop

The next stage is to complete Orislop’s multimodal detection system by combining temporal, spatial, audio, audio-video synchronization, repost-detection, and content-provenance signals.

We also plan to:

Expand and manually verify the training dataset
Improve classification accuracy and reduce false positives
Add user feedback that helps personalize and retrain the system
Optimize lightweight local models for faster private inference
Use cloud analysis only when more computation is required
Strengthen support for YouTube Shorts, TikTok, Instagram Reels, and LinkedIn
Release the extension to an initial group of real users
Develop mobile versions for Android and eventually iOS
Create tools that allow users to understand and customize exactly why content is filtered

Our long-term goal is not simply to build a slop detector. We want to create a user-controlled intelligence layer that sits between people and algorithmic feeds, giving users the ability to decide what kinds of content earn their attention.