Dialectic

Generic AI summarizers notoriously distort debate nuances: they fabricate non-existent quotes, smooth over genuine disagreements, or misattribute arguments to the wrong speaker. Meanwhile, rendering real debates with branching rebuttals, concessions, and joint-support junctions quickly becomes an unreadable tangle of visual chaos.

The engineering mandate demanded absolute citation integrity, sub-50ms local responsiveness, and an intuitive canvas capable of organizing up to 160 intricate argument nodes without freezing the browser.

• Guaranteeing that every single extracted quotation is strictly grounded in exact, verifiable character offsets of the source transcript.
• Architecting an automated ELK layout engine that automatically handles circular reasoning, shared evidence, and multi-premise junctions.
• Eliminating model bias by evaluating exchanges across three orthogonal criteria: formal logic, evidence strength, and conversational responsiveness.

Rather than positioning AI as an opaque judge of truth, Dialectic treats it as a structured analytical assistant whose every output is inspectable by the user. Every card links directly to its source quote, timestamp, and speaker attribution.

We designed a multi-stage pipeline supporting both Google Gemini for rapid zero-cost extraction and OpenAI Responses for structured synthesis with on-demand web research. Everything is stored locally on the user’s machine in an atomic file database, ensuring full privacy and zero vendor lock-in.

• Multi-pass extraction pipeline dividing speaker diarization, topic chunking, connection reconciliation, and exchange adjudication.
• On-demand web research for individual factual claims with strict citation retention, downgrading ungrounded claims to unverified.
• Local-first atomic persistence in .data/ protecting user notes, custom spatial arrangements, and review flags across sessions.

We engineered the web client with React 19, TypeScript, and @xyflow/react, creating a fluid infinite canvas with custom argument cards, collapsible topic clusters, and responsive layout algorithms. Custom node types render logical relationships — support, challenge, qualification, and question — with distinct visual grammar.

The backend runs on Node.js and Express 5, implementing strict Zod schema validation across all pipeline stages. The server automatically rejects fabricated quotations before they can reach the UI, while a debounced, serialized workspace saver protects user canvas adjustments.

• Integrated media caption ingest supporting YouTube, Vimeo, and TED via yt-dlp, alongside raw TXT, SRT, and VTT transcript uploads.
• Sub-50ms canvas interaction latency with smooth hardware-accelerated zoom, pan, and branch collapse.
• Dual provider flexibility allowing seamless toggling between Gemini 3.8 Flash and OpenAI gpt-4.1 with structured outputs.