ContextBridge AI

Inspiration

Learning from long technical videos is often passive. Learners may understand better in their own language, struggle to find the exact moment where something was explained, or ask questions that the video simply doesn't cover.

We wanted to turn a video from something you watch into something you can question, verify, and explore.

What it does

ContextBridge turns a video into an interactive learning workspace.

Ask questions in English, Hindi, or Hinglish Get answers grounded strictly in the selected video Jump directly to the exact timestamp and transcript evidence Clearly identify when an answer is not covered by the video Search the web separately when the learner chooses to go beyond the video Detect contradictions within the lesson Suggest relevant concepts to explore next Support voice-based interaction and read-aloud answers How we built it

Built with Google Gemini 2.5 Flash, Google Search Grounding, FastAPI, Next.js 16, and Google Cloud Run.

The video is analyzed into a structured transcript, chapters, topics, and events. For each question, ContextBridge retrieves relevant video context, generates an answer, and validates the answer against the transcript before showing it.

The system uses a three-tool agent loop with strict evidence validation. Video answers require matching transcript quotes and valid timestamps; unsupported questions are explicitly marked instead of being fabricated.

Challenges we ran into

The hardest challenge was trust.

A normal LLM can produce a convincing answer even when the video never said it. We had to design the system so that an answer cannot simply sound correct — it must be backed by the video's actual transcript and timestamp.

We also had to handle language switching, video synchronization, external research boundaries, and graceful failure when AI or external services are unavailable.

Accomplishments that we're proud of

0% unsupported-answer rate in our 25-case evaluation 100% timestamp retrieval accuracy in the latest local benchmark 100% groundedness with transcript-verified quotes Exact timestamp → video + transcript synchronization Hindi, English and Hinglish interaction Contradiction detection Explicit boundary between video knowledge and web research 83/83 tests passing in the repository

What we learned

We learned that building a reliable AI product is not just about choosing a powerful model.

The important part is designing boundaries, evidence, validation, and failure handling around the model so users know not only what the AI says, but why they should trust it.

What's next for ContextBridge

Expand multilingual support beyond Hindi and English Support more educational content formats Improve personalization and learning paths Add richer learner progress and knowledge tracking Scale the evidence-grounded architecture for larger video libraries

Built With

• agentic-ai
• cloud-run
• edtech
• education
• fastapi
• generative-ai
• google-gemini
• google-search-grounding
• google-vertex-ai
• multimodal-ai
• next.js
• python
• react
• speech-ai
• typescript