Attune

Framing the Problem

Conference libraries contain hours of valuable content, but learners often struggle to identify which talks match their experience and goals. Even after watching, key lessons are easily forgotten. This especially affects self-directed and introverted learners who want guidance without judgment or information overload.

Idea Explanation

Attune is a personalized watching companion that helps learners watch what fits and keep what matters.

Users choose their experience level and interests, and Attune ranks conference talks by personal fit rather than perceived quality. While a talk plays, distilled claims, action items, and resources appear in sync with the speaker. Users can also ask questions and receive transcript-grounded explanations adapted to their level.

Implementation

Attune is built with Next.js, React, TypeScript, and KendoReact.

The frontend handles onboarding, deterministic talk ranking, video playback, and synchronized note reveals. Talk metadata, transcripts, fit facets, and pre-generated distillations are stored as validated JSON files, so the project does not require a database.

Most features work without network requests. Only the “Explain that” chat calls a Next.js backend API route, which securely sends the relevant transcript context to Gemini and returns a level-appropriate answer. Zod validates all structured project data, while Vitest verifies ranking behavior and data integrity.

Challenges

The main challenge was using AI without making the demo slow, expensive, or unreliable. I solved this by generating talk facets and distillations ahead of time, then using a deterministic scoring function in the browser for personalization.

Another challenge was synchronizing AI-generated notes with real video content. Instead of asking the model to invent timestamps, each note references a known talk section. The application resolves that section to its verified timestamp, making every insight traceable to the original recording.

I also had to frame recommendations carefully. Attune judges personal fit, never speaker or talk quality, so every recommendation remains constructive.

Accomplishments

I built an end-to-end experience that turns a static conference archive into a personalized learning journey.

I learned how to combine deterministic software with generative AI: code handles ranking, timing, and validation, while AI handles language understanding and explanation. I also created a resilient architecture where the core experience works without an API key, and live AI is reserved for the feature where it adds the most value.

Next Steps

Next, I would add user accounts, saved learning libraries, progress tracking, and recommendations across multiple conferences. I would also support more video platforms, multilingual transcripts, searchable notes, and exports to tools such as Notion or Markdown.

Longer term, Attune could help conference organizers make their full content catalog more discoverable while giving learners a lasting, personalized record of what they watched.