Backbencher is a Chrome extension built for students who learn from video lectures. Instead of scrubbing through a 90-minute recording to find a specific concept, you can ask Backbencher a question and get a direct answer — grounded in what the lecturer actually said.
What it does
The extension works in two modes:
- Summarise — Generates a structured summary of the lecture, broken down by topic. Great for quick revision or deciding whether a particular lecture is worth watching in full.
- Chat — Ask anything about the video. “What did the lecturer say about attention mechanisms?” or “Summarise the part about gradient descent” — Backbencher retrieves the relevant transcript segments and generates a response from them.
Both modes work without ever sending the full transcript to the model on every query.
The RAG pipeline
The core challenge was context length. A 90-minute lecture transcript can be 15,000+ tokens — well beyond what you’d want to stuff into every prompt, both for cost and latency reasons.
The solution was a retrieval-augmented generation pipeline:
- The transcript is chunked and embedded on the client.
- The user’s question is embedded and used to retrieve the top-k most relevant chunks via cosine similarity.
- Only those chunks are passed to the LLM, keeping the prompt tight and the response fast.
This meant the per-query token cost dropped significantly compared to naive full-transcript prompting.
Multi-agent token optimisation
The bigger architectural challenge was storage and retrieval across sessions. Embedding and storing a full transcript per video, per user, per session was expensive at scale.
The solution was a multi-agent system where a coordinator agent routes incoming requests to specialised sub-agents:
- A transcript agent handles chunking, embedding, and caching — so the same lecture isn’t re-processed if another user has already watched it.
- A retrieval agent handles similarity search and context assembly.
- A generation agent handles the final prompt construction and LLM call.
This separation made it easy to swap models, tune retrieval independently of generation, and keep token usage auditable per agent.
Beyond engineering
Backbencher wasn’t just a technical build — I also handled the marketing and brand design: the visual identity, the landing page, and the growth strategy. Working across both sides of the product gave me a much clearer picture of how feature decisions affect user perception, and vice versa.