ScrollSmart

ScrollSmart Report

1. Project Inspiration

ScrollSmart was inspired by a simple but important observation: many people spend a large amount of time doomscrolling short-form content, but often finish that experience feeling empty, overstimulated, and intellectually unsatisfied. The emotional reward is immediate, but the long-term value is low. We wanted to keep the frictionless, addictive, swipe-based interaction that makes short-form video platforms so compelling, while replacing low-value content with personalized knowledge experiences.

Our core idea was to transform scrolling from a passive habit into an active learning loop. Instead of endless entertainment that disappears without impact, ScrollSmart turns each swipe into an opportunity to discover new ideas, deepen curiosity, and better understand one’s own interests. In that sense, the product is not trying to fight the scroll behavior directly; it is trying to redirect it toward something more meaningful.

This matters especially in the era of generative AI. Today, most AI systems are reactive: users must already know what to ask. In reality, many users do not know what they are curious about, how to phrase a good question, or where to begin. ScrollSmart addresses that gap. The system proactively starts the learning interaction, gives users something interesting to react to, and gradually teaches them what to ask next. That shift, from “asking AI a question” to “learning how to ask better questions through AI,” is one of the most valuable educational opportunities created by generative AI.

2. Product Summary

ScrollSmart is a TikTok-style AI knowledge feed. Users first choose topics they are interested in, then scroll through personalized knowledge cards generated by AI. Each card begins with a strong “hook” designed to spark curiosity, and users can then:

like or skip the content,
click “Go Deeper” for more advanced explanation,
ask follow-up questions in a chat interface,
or simply spend more time on topics they find engaging.

As users interact, the system updates an evolving interest profile. That profile is visualized inside the product through a live radar chart and topic-weight display, so users can see how their intellectual preferences develop over time. The product therefore serves two purposes at once: it delivers knowledge, and it helps users discover who they are intellectually.

3. Core Product Functions

The main functional components of ScrollSmart are:

Topic onboarding: users select interest areas at the beginning.
Personalized knowledge feed: the system generates the next card based on the user’s evolving profile.
Interactive learning cards: each card supports likes, skips, “Go Deeper,” and free-form follow-up questions.
Live profile visualization: the user can view how the system interprets their interests through topic weights and radar visualization.
Continuous adaptation: every interaction influences future recommendations.

This creates a loop of: exposure, reaction, adaptation, and refinement.

4. Technical Architecture and Stack

The project uses a modern full-stack architecture:

Frontend: Next.js 14, TypeScript, Tailwind CSS, Framer Motion, Recharts
Backend: FastAPI, Python
AI model layer: OpenAI API
Development platforms and AI tooling: GitHub, Cursor, Codex, Gemini, ChatGPT, and other AI-assisted tools

More than 70% of the development workflow involved AI-assisted coding, iteration, debugging, or design support. However, the product logic itself is not just “AI generating everything blindly.” The important system behavior is structured through explicit backend logic, which is one of the project’s technical strengths.

5. Backend Algorithm and Why It Is Strong

The backend uses a dual-agent design with strict separation of responsibilities:

Agent 2: Analyst
- updates the user’s interest profile,
- chooses the next topic,
- writes a natural-language summary of the user’s interests for personalization.
Agent 1: Content Generator
- creates the hook message for the next card,
- produces deeper explanations,
- handles ongoing card-level conversation.

This separation is important because it avoids role confusion. The recommendation logic is not mixed together with creative text generation. That reduces prompt contamination, makes the system easier to debug, and creates a cleaner architecture overall.

The recommendation engine uses three important mechanisms:

A. Explicit interest weights Each topic has a numerical weight. These weights are updated using interpretable behavioral signals:

Like: +0.15
Skip/Dislike: -0.10
Go Deeper: +0.20
Each chat message: +0.05, capped at +0.30
Long dwell time over 30 seconds: +0.10

This is a strong design choice because it prevents the system from becoming a black box. Instead of relying entirely on vague AI intuition, the product uses measurable signals tied to meaningful learning behavior.

B. 80/20 recommendation strategy The next topic is selected using an 80/20 exploitation-exploration rule:

80% of the time, the system selects from higher-weight topics
20% of the time, it explores lower-weight or less-seen topics

This avoids two common recommendation failures:

overfitting too quickly to a narrow interest cluster
becoming repetitive and boring

In other words, the user still gets a personalized experience, but the feed remains capable of discovery.

C. Soft normalization Weights are constrained and softly normalized over time. This prevents all topics from drifting upward uncontrollably and avoids a situation where every topic becomes equally “important.” Without this safeguard, the profile would lose meaning and recommendations would become noisy.

Together, these design choices make the backend logic robust, interpretable, and better suited for educational personalization than a purely generative feed.

6. Why the System Improves Educational UX

ScrollSmart combines the strongest engagement mechanics of short-form media with the strongest strengths of AI tutoring:

low friction,
instant novelty,
adaptive personalization,
and conversational depth.

This creates an experience that feels smooth and familiar, but leads to learning rather than emptiness. Users reported that the feed feels very fluid and natural, with a strong sense of “real scrolling,” similar to the interaction style of apps like TikTok. At the same time, unlike traditional doomscrolling, the experience leaves users with useful knowledge and a stronger sense of curiosity.

Another key benefit is that ScrollSmart lowers the barrier to intellectual exploration. Users do not need to enter the app already knowing what they want to study. The system introduces surprising ideas first, and then lets the user deepen selectively. This makes the educational journey feel playful rather than demanding.

7. User Experience Feedback

Early user impressions can be summarized as follows:

The scrolling experience feels smooth, polished, and highly responsive.
The interaction pattern feels familiar, making the product easy to adopt immediately.
Users feel they are actually gaining knowledge while using the app.
The product combines the dopamine-driven strengths of short-form content with meaningful educational value.
The evolving profile visualization helps users reflect on their own interests, which adds a layer of self-discovery beyond simple content consumption.

This is an important UX achievement: the product does not ask users to sacrifice engagement in order to learn. Instead, it makes learning itself feel engaging.

8. Why This Matters

ScrollSmart points toward a broader future for generative AI in education. Rather than replacing teachers or acting only as a passive chatbot, AI can become a dynamic curiosity engine: something that introduces ideas, adapts to behavior, scaffolds exploration, and helps users form better questions over time.

In that sense, ScrollSmart is not just a content app. It is a new interaction model for AI-powered learning: one where entertainment mechanics and educational value are no longer opposites, but can reinforce one another.

If you want, I can turn this into: