AI Breakfast Shanghai
← Back to Notes

AI Breakfast #18

Executive Summary

At our AI Breakfast, our group of developers, entrepreneurs, educators, and data scientists discussed topics ranging from AI tools for reading and pronunciation, challenges in verifying information with AI, to OCR solutions for complex documents. The attendees also shared the latest on their work and projects, including an AI agent for speech assessment and a German language learning app.

Member Work

AI Speech Assessment Agent

A new member introduced their company, Communication Intelligence Labs, which builds AI agents to assess speaking ability for students. Drawing from a background in competitive debate, the founder and their team created a system that analyzes video recordings of student presentations. The tool breaks down performance across 28 skill markers, including content organization, evidence usage, voice modulation, and body language.

The group discussed the technical implementation, which involves a multi-agent system:

  • Content Analysis: Uses Large Language Models (LLMs) like Gemini to evaluate the script based on a proprietary rubric (e.g., identifying a "central idea" or "evidence").
  • Visual & Audio Analysis: Tracks gestures, eye contact, and vocal pitch. The founder noted challenges in quantifying subjective metrics like "good" eye contact or appropriate volume, but highlighted that gesture tracking is already proving effective.
  • Target Market: Currently focused on K-12 science presentations in international and bilingual schools, where standardized behaviors make assessment more reliable than in general public speaking.

Lumi: German Reading Assistant

The founder of Human Meets AI shared updates on "Lumi," an app designed to help children learn to read German. The app addresses the challenge of parents not always being available to correct their child's reading.

  • Functionality: The child reads a story aloud, and the app uses speech-to-text technology to analyze pronunciation, speed, and accuracy. It highlights specific words that were mispronounced or skipped.
  • Technology: It leverages Microsoft's speech services for pronunciation assessment (noted by the group as unique for providing phoneme-level feedback) and generates story illustrations using the "Nano Banana" image generation model.
  • User Experience: The group discussed the importance of vocalization in learning to read, noting that speaking aloud helps internalize correct pronunciation and prevents the brain from "skipping" over complex grammar structures common in German.

Group Discussions

Gratitude for AI Tools

In the spirit of Thanksgiving, the group shared which AI technologies they are most grateful for:

  • Voice-to-Text: Multiple members praised the reliability of modern dictation and translation tools, which have become essential for daily tasks like setting alarms or drafting text.
  • Coding Assistants: Tools like GitHub Copilot and Cursor's Composer model were highlighted as game-changers, enabling non-coders to build apps and helping experienced developers work faster.
  • Image Generation: Members expressed appreciation for tools that allow those without artistic skills to visualize ideas, specifically mentioning the "Nano Banana Pro" model for its high-quality outputs compared to older models.

OCR & Document Processing

The conversation turned to the difficulties of extracting data from PDFs, particularly those containing a mix of text, images, and charts.

  • The "PDF Problem": Members noted that PDFs are designed for visual consistency, not data extraction, making them difficult to parse.
  • Solution Spotlight: A member shared their success with Dots OCR, an open-source tool from Xiaohongshu. It excels at identifying the coordinates of different elements on a page—segmenting text, images, and LaTeX formulas so they can be processed individually. While it can be slow (taking up to 50 seconds per page in some hosted environments), it provides high accuracy for complex documents.

Information Validity: Wikipedia vs. AI

A debate emerged regarding the reliability of AI-generated knowledge bases (like Grok's "Rockopedia") versus human-curated platforms like Wikipedia.

  • Human vs. Machine: One member, a history teacher, defended Wikipedia, noting that its human editors are often passionate experts dedicated to accuracy. They expressed skepticism that an AI could replicate the nuanced editorial policies and conflict resolution processes of human communities, especially on controversial topics.
  • The Scale Advantage: Countering this, another member argued that AI's true power lies in scaling effort. Just as AlphaFold solved protein folding by scaling computation, AI could potentially process vast physical archives and niche topics that human editors haven't reached, democratizing access to information that currently sits in basements or untranslated texts.

Tech Infrastructure & Global Perspectives

The group touched on the unique challenges and observations of working in tech across different regions:

  • The "China Tax": A member coined this term to describe the time cost paid by developers in China to manage VPNs and proxies just to access standard development tools like Docker, Claude, or Google AI Studio.
  • China's Tech Progress: There was a discussion on the disconnect between Western perceptions and the reality of China's technological advancement, particularly in robotics and electric vehicles (EVs). Members observed that while some demos are dismissed as "fake" abroad, the on-the-ground reality often shows functional, advanced hardware that rivals or exceeds international competitors.

Other Resources

  • Dots OCR: An open-source OCR tool from Xiaohongshu that effectively segments PDFs into text, images, and formulas.
  • Vercel v0: A generative UI system that allows users to create website interfaces via text prompts; used by members to generate design schemes.
  • Cursor (Composer Model): An AI code editor feature that allows for multi-file editing and faster development workflows.
  • Vertex AI: Google's machine learning platform, used by members to access various models for image generation and processing.
  • Rockopedia: A feature referenced in relation to Grok, attempting to build a knowledge base similar to Wikipedia using AI.
← Back to Notes