Marketplace

Confidential Meeting Notes Summarizer

The Confidential Meeting Notes Summarizer is a two-stage, privacy-first AI agent that processes meeting notes or transcripts entirely inside the browser — no data ever leaves your device. What makes it technically unique is the anonymization pipeline that runs before the language model ever sees your text. In stage one, a local BERT-based Named Entity Recognition (NER) model — onnx-community/bert-base-NER-ONNX running via Transformers.js — scans your raw notes and detects persons (PER), organisations (ORG), and locations (LOC) with confidence scores above 0.85. Every detected entity is replaced with a neutral label like [PER-1], [ORG-2], or [LOC-3] before the text is passed anywhere else. An anonymization report shows you exactly what was replaced, what it was replaced with, and the entity type — so the process is fully transparent and auditable. In stage two, four Python tools running via Pyodide perform deterministic pre-analysis on the anonymized text: action item extraction (patterns like "Action:", "TODO:", "will", "needs to", "assigned to"), decision detection ("decided", "agreed", "approved", "we will"), topic scoring across eight categories (budget, timeline, technical, product, HR, sales, strategy, risks), and meeting metadata extraction (word count, estimated duration, transcript vs. bullet format, participant count). These structured results are passed as a context block to a fully local WebLLM language model (Hermes-2-Pro-Mistral-7B via WebGPU), which generates a four-section prose output: Summary, Decisions, Key Discussions, and Tone & Dynamics. Because the entire pipeline — BERT NER, Python analysis, and LLM inference — runs locally in the browser, this template is suitable for legal proceedings, HR interviews, board meetings, client calls, medical consultations, or any meeting where confidentiality is non-negotiable. No API key required. Works completely offline after initial model download.