Inspiration
Long-form YouTube videos contain valuable knowledge, but finding one specific idea again often means searching through hours of footage. We wanted a private research tool that could collect available captions, preserve their sources, and make the resulting knowledge searchable without requiring AI for every query.
The caption-retrieval foundation was inspired by the freely shared Transcribeer skill from Bart Boonstra of Slim Werken AI, whose attribution has been retained.
What it does
Transcript Knowledge Hub is a local-first Windows application that monitors selected public YouTube channels and retrieves available captions without downloading video or audio.
Transcripts are stored locally as readable Markdown files. Users can search their entire archive without an AI service or API key. They can also select trusted transcripts and use GPT-5.6 Terra to generate structured Knowledge Briefs containing summaries, key insights, evidence, uncertainties, next steps, and a source map.
The application supports Dutch and English, with separate controls for the interface language, transcript preference, and AI output language. Before transcript text is sent to OpenAI, the user must explicitly confirm the transmission.
How we built it
We built the desktop interface with Python and Tkinter. SQLite stores channels, video metadata, queue state, settings, and analysis history, while Markdown files provide a durable and portable transcript archive.
The application uses yt-dlp for public channel and video metadata, youtube-transcript-api for available caption tracks, and the OpenAI Responses API for Knowledge Briefs. GPT-5.6 Terra is the default analysis model, and API keys are stored through Windows Credential Manager.
Codex helped us inspect the architecture, implement localization, develop the analysis workflow, strengthen error handling, create automated tests, build a synthetic demonstration mode, verify the release package, and prepare the submission documentation.
Challenges we ran into
The first major challenge was separating three concepts that initially looked like one setting: interface language, transcript language preference, and generated-output language. Each required different behavior and persistence.
YouTube caption availability and throttling created another challenge. We needed durable queue states, retries, cooldowns, and clear status messages instead of assuming that every public video would have an immediately accessible transcript.
We also needed to prepare a demonstration without exposing an operational database, personal file paths, API keys, or real transcript content. We created an isolated synthetic demo that disables background activity and removes its temporary data when closed.
Finally, we had to continue development without modifying or restarting the production instance that was already running.
Accomplishments that we're proud of
We created a complete English and Dutch workflow, separate language controls, local full-text search, source-linked Knowledge Briefs, explicit privacy confirmation, secure credential storage, and a clean synthetic demonstration.
The candidate passed 33 automated tests, application and GUI self-tests, privacy and secret scans, clean-archive verification, and a successful installation test on a second Windows PC.
We are especially proud that normal transcript search remains completely local and does not require an API key or AI request.
What we learned
We learned that local-first design improves more than privacy. It also makes results easier to inspect, trace, export, and verify.
We learned that language support requires clear product semantics rather than simply translating labels. We also discovered that packaging, attribution, testing, and honest provenance are as important as visible features.
Finally, we learned that AI is most useful when it remains an explicit, traceable step instead of becoming an invisible dependency throughout the application.
What's next for Transcript Knowledge Hub
Next, we want to add richer knowledge profiles that combine selected channels and transcripts into reusable, source-linked research collections.
We also plan to improve contradiction detection, timeline extraction, reusable analysis templates, and exports for Codex skills and other knowledge-management tools.
Additional priorities include broader Windows testing, improved onboarding, more granular source citations, and continued refinement of the multilingual workflow.
Log in or sign up for Devpost to join the conversation.