Transcript Intelligence Pipeline

A clean-room Python proof of concept that converts noisy conversations into privacy-safe, decision-ready evidence.

Pipeline

1. Remove IVR, system, and bot noise
2. Normalize speaker turns
3. Redact email, phone, and long numeric identifiers
4. Classify operational issues
5. Build issue fingerprints for repeat contact, ownership, timing, handoff, and customer effort
6. Select concise evidence excerpts
7. Produce a recommended operational action and executive brief

Improvements demonstrated

• Keeps the normalized transcript and the executive narrative separate
• Adds explicit quality checks for speaker labels, customer evidence, and classification
• Limits evidence excerpts to short, reviewable passages
• Creates issue fingerprints rather than relying only on broad topic labels
• Preserves uncertainty by marking unmatched conversations as unclassified

Run

python run_demo.py
python -m pytest -q

Open outputs/executive_brief.html.

Clean-room notice

All transcripts, speakers, issues, evidence excerpts, scores, and findings are fictional and independently created. No employer data, customer data, source transcripts, private prompts, or confidential metrics are included.

Scope

This demonstration uses transparent lexical classification. A production system would require approved data access, stronger privacy review, taxonomy governance, human quality assurance, and appropriately validated language models.