Perchspective

Inspiration

From 1994 to 2014, the percentage of Americans identifying with either extreme of the political spectrum doubled. In the past decade, this polarization has been further amplified by content bubbles that provide an echo chamber for confirmation bias. Video platforms like YouTube and Tiktok often prioritize engagement-driven algorithms that surface content aligned with users' existing beliefs, reinforcing a single perspective rather than encouraging exploration. We created Perchspective to empower them to explore multiple viewpoints on the same topic and engage with information more thoughtfully.

What it does

Perchspective lets users input a video link and automatically analyzes the content across several dimensions. It summarizes the topics covered, provides perspective analysis—where the content falls on the political spectrum, its emotional tone, factual density, source credibility, and diversity of viewpoints. The platform also suggests alternative perspectives on the same topic and provides links to related videos, helping users broaden the range of information they consume. By providing this insight, Perchspective not only makes users more aware of the content they're viewing but also gives them control over their exposure, helping them engage with multiple viewpoints in an effort towards reducing political polarization.

How we built it

We built Perchspective using a streamlined pipeline. First, we extract the transcript from a video (currently supported for YouTube Shorts that already have captions available on the platform). We then pass the transcript along with the video title to a pretrained LLM, Gemini 2.5 Flash, which generates a structured JSON output containing the required analytical dimensions–including topic, perspective analysis, and alternative viewpoints. This structured output is then processed and presented clearly on our website for the user.

Challenges we ran into

We faced several technical constraints while building Perchspective. Our use of the free version of Gemini limited request volume and model flexibility. Additionally, not all videos provide publicly available transcripts, which restricted the range of content we could analyze. Without GPU access, we were unable to fine-tune models on labeled data to improve prediction accuracy or customize outputs for more nuanced perspective detection. These limitations required us to rely heavily on prompt engineering and efficient API usage to achieve reliable results.

Accomplishments that we're proud of

We’re proud of how Perchspective pairs a clean, intuitive UI with analysis that is grounded in the video transcript itself, rather than inferred from the title or channel name. The platform doesn’t just present classifications — it supports them with clear reasoning, highlighting why a particular bias, tone, or framing was detected. Achieving this required a carefully crafted prompting strategy, giving us hands-on experience with prompt engineering to guide the AI into producing structured, insightful, and reliable outputs.

What we learned

We gained practical full-stack experience by connecting our backend logic — including transcript processing, LLM API calls, and structured outputs — with the frontend interface of our website. Along the way, we also learned how to use AI tools effectively to streamline development and generate more accurate, structured results.

What's next for Perchspective

Our next step is to add browser extension capability to Perchspective, allowing it to work directly on YouTube and provide real-time recommendations. Instead of requiring users to paste a link into a separate website, the extension would integrate seamlessly into the viewing experience, enabling analysis and alternative perspective suggestions with a single click. To expand the range of videos we can analyze, we aim to integrate automatic transcription models like Whisper so we can process videos without requiring existing captions. We want Perchspective to move beyond isolated, one-time analysis and give users a better understanding of their overall content exposure. In future versions, each video’s political-leaning score would be placed in context by storing past analyses — including video ID, title, thumbnail, leaning score, tone score, and source domain — and generating a histogram view. This would allow users to see how a video compares to what they’ve watched previously and to broader viewing patterns across the platform, helping them understand how their feed functions and whether they are consistently exposed to similar viewpoints or engaging with diverse perspectives.

Built With

• css
• fastapi
• gemini
• javascript
• next.js
• node.js
• python
• react
• tailwind
• typescript