Documents you can talk to — with citations

BanterBox lets readers ask a document questions, interview its author, or talk to a story's characters — every answer grounded in the source text with tap-to-verify citations. It spans three reading experiences — research & policy, education, and fiction — and gives content owners an on-brand, analytics-rich conversational layer over their own material.

Long-form work — strategy reports, policy briefs, textbooks, novels — increasingly gets skimmed or pasted into general chatbots, where answers are ungrounded, uncited, and disconnected from the author's intent. Publishers and institutions lose control of how their work is read, and get no insight into engagement. BanterBox was built so content owners can offer their own audience an on-brand, source-cited conversational layer — and finally see how readers actually engage.

We built a Python/Flask platform around a tiered retrieval engine — vector embeddings, BM25 keyword search, and a cross-encoder reranker — that routes questions through a multi-provider LLM layer (Claude Sonnet as the default, with OpenAI, Google, and Perplexity fallbacks) and returns every answer with tap-to-verify citations back to the source passage. An ingestion pipeline turns uploaded PDFs into a navigable 'argument tree,' and supporting services add text-to-speech narration, real-time translation, and a content-repurposing studio. The reading experiences ship as React/TypeScript apps over server-rendered portal pages, deployed on Render with Postgres, Redis, Celery, and AWS S3 + CloudFront, behind Auth0 and white-label client domains.