albert-ying/agentic-cortex/detect-people

Detect People

Create user entries for people detected in screenpipe data. Uses AI judgment to filter false positives.

Usage

# Run candidate extraction
python3 scripts/extract_candidate_names.py --date 2026-03-14
 
# Then validate candidates and create entries (agent step)

4-Step Process

1. Extract candidates

python3 scripts/extract_candidate_names.py --date YYYY-MM-DD

Returns JSON with OCR/audio candidates and existing user slugs.

2. Validate with AI judgment

For each candidate, determine if it's a real person:

• Create entry -- real person with professional context
• Link to existing -- audio name matches existing user phonetically
• Skip -- UI element, device name, OCR artifact

Skip rules -- these are NOT people:

• UI elements: "App Store", "Enhanced Context", "System Preferences"
• Device names: "HyperX SoloCast", "MacBook Pro"
• OCR artifacts: "Starting VideoCapture", "Screenshot Taken"
• Generic labels: "Unknown User", "Admin Panel"

When in doubt, skip. A missed person can be added later; a false entry pollutes the vault.

3. Create entries with web search

For each validated person:

1. Web search: name + "professor" OR name + "researcher" OR name + institution
2. Extract: institution, role, email (if public), bio summary
3. Create entry with ALL fields filled -- never use placeholders like "[To be filled]"

If web search finds nothing: skip creating the entry (likely false positive).

Generate a 23-character lowercase alphanumeric node ID for each entry:

import random, string, time
node_id = ''.join(random.choice(string.ascii_lowercase + string.digits) for _ in range(23))
timestamp_ms = int(time.time() * 1000)

Entry template (user.<name-slug>.md):

---
id: <23-char-lowercase-alphanumeric>
title: Full Name
desc: 'One-line role/institution'
updated: <unix-ms>
created: <unix-ms>
---
 
## Contact Info
 
-Email: actual-email@example.com
-Institution: University / Company
-Web: https://profile-url
 
## Role & Expertise
 
Title and role description.
 
Brief research or professional summary.
 
## Context
 
**How we connected**: Detected from screenpipe journal
 
**Date**: YYYY-MM-DD
 
**Context**: Description of how they appeared (meeting, email, mention)
 
## Notes
 
Any relevant observations from the detection context.

4. Return detected people as JSON

Return results for the calling skill (e.g., daily-journal) to use for linking:

{
  "people": [
    {"name": "Sarah Kim", "slug": "sarah-kim", "status": "existing"},
    {"name": "David Lee", "slug": "david-lee", "status": "created"}
  ],
  "summary": {
    "created": 1,
    "linked": 1,
    "skipped": 35
  }
}

Brief text summary:

• Created: N entries (list names)
• Linked: N audio names to existing users
• Skipped: N false positives (total only)

Key Principle: When in doubt, skip. All created entries must be fully populated with real information from web search. No placeholders.

Auto-run: This skill is automatically called during the daily-journal workflow.