Yuh Hear Dem?

Yuh Hear Dem? turns hours of Barbados parliament videos into searchable knowledge graphs—so citizens can explore who said what, when, and why, without watching the whole thing.

Comment

🧠 Inspiration

We’re a father–daughter team — Matt (in Barbados) and Katie (in the UK) — building this together over the past few weeks and submitting it on Father’s Day!
This project started as a way for Katie to learn more about AI and large language models, with Matt mentoring on backend and architecture.

But it quickly turned into something bigger: a tool to make Barbados Parliament videos searchable, transparent, and easier for people to engage with — especially those who don’t have the time to scrub through hours of footage.

We wanted to answer a simple question:
"What did they really say?"
And make that answer accessible to anyone in the country.

🔍 What It Does

YuhHearDem takes long, unstructured YouTube recordings from the Barbados Parliament and turns them into a conversational knowledge assistant.

You can ask questions like:

🏥 “What did the Minister of Health say about the sugar tax?”

And YuhHearDem will:

  • Search a knowledge graph of topics and entities
  • Find the exact moment in a video where it was said
  • Provide timestamped links to go straight to the source
  • Suggest follow-up questions using graph-based context

It's chat-first civic intelligence powered by LLMs and big data.

🛠️ How We Built It

  • We ingest YouTube transcripts from parliamentary sessions
  • Clean and align them using Gemini Flash
  • Extract entities, topics, and relationships into a knowledge graph
  • Store graph data and vector embeddings in MongoDB Atlas
  • Run hybrid GraphRAG search at query time (graph + vector)
  • Serve responses through a chat UI powered by Google ADK

Matt focused on the backend, knowledge representation, and retrieval pipeline.
Katie designed and built the frontend UX — focusing on accessibility and making the interaction feel intuitive and natural.

🧗 Challenges We Ran Into

  • Transcript cleanup: YouTube captions are messy — full of stutters, cutoffs, and misalignment.
  • Graph consistency: Making sure every node was connected, searchable, and meaningful.
  • Async tooling: Getting Gemini, ADK, MongoDB, and the frontend to talk reliably in real time.
  • Timezones & time limits: Juggling day jobs, time differences, and still getting this live!

🏅 Accomplishments That We're Proud Of

  • Built an end-to-end pipeline from video to knowledge graph to conversational AI
  • Created a system that works at scale and can be reused for other civic datasets
  • Learned new tools (Gemini, MongoDB Atlas, ADK) and made them work together
  • Watched Katie go from frontend learner to shipping a production-grade AI interface

And of course — submitting this as a family project on Father's Day ❤️

📚 What We Learned

  • How to design structured knowledge graphs from messy real-world transcripts
  • How to build LLM-powered tools that are actually grounded in data
  • How to make chat interfaces that don’t feel like tech demos, but real tools
  • That big data + small teams can build real civic change

🚀 What’s Next for YuhHearDem?

  • Expand to cover the Prime Minister’s Office channel and statutory legislation
  • Let users track politicians’ stances across sessions and topics
  • Add summarization, alerts, and a public mobile version
  • Work with journalists and civic groups to make this even more impactful

This is just the beginning.

Share this project:

Updates