predictleads-dashboard

PredictLeads Dashboard (HTML viz)

Generates a single self-contained HTML page from cached signals in ~/.gtm-os/gtm-os.db. Cards per company with signal-count badges, top-signal callout, expandable detail (recent jobs, news, funding, tech stack, similar companies). Filter by vertical, sort by signal density or recency. Auto dark/light. Zero API calls.

When to use

• After running prospect-discovery-pipeline to scan all 10 finalists in one view
• After bulk-enriching a campaign result set (signals:enrich --result-set) for a visual sanity check before outreach
• Sharing signal context with a non-technical teammate (open the HTML, no CLI knowledge needed)

Don't use when: you only have signals for 1–2 companies (just use signals:show); signals haven't been pulled yet (run signals:fetch first).

How to invoke

The dashboard is built by a small Python script. Pass a list of domains and an optional list of pre-built lead cards (name + title + LinkedIn URL).

Inputs the skill needs

1. List of domains (must already be in company_signals table)
2. Optional per-domain lead metadata: { company, vertical, geo, lead_name, lead_title, linkedin }

Build steps

1. Read the lead metadata into a Python dict (see existing template at ~/Desktop/predictleads-dashboard.html for shape).
2. Query SQLite for each domain:
3. SELECT signal_type, COUNT(*) for badge counts
4. Top 8 jobs by event_date DESC
5. Top 8 news by event_date DESC
6. Top 5 financing events
7. Top 12 technologies
8. Top 10 similar_companies sorted by payload.score
9. Render the HTML template (see Implementation below) with embedded JSON.
10. Write to ~/Desktop/predictleads-dashboard-{client_or_topic}-{date}.html and open it.

Implementation

A Python generator script lives at scripts/predictleads-dashboard.py (when committed). It reads from ~/.gtm-os/gtm-os.db, accepts a JSON config of leads, and emits a self-contained HTML file.

If the script is missing, model the new one on the prior run captured at ~/Desktop/predictleads-dashboard.html (Apr 30 2026). Key visual elements to keep:

• Per-company card with company name + vertical tag (color-coded) + domain
• Marketing lead pinned at top of each card with LinkedIn link
• 5 signal-type badges with counts (jobs / funding / news / tech / similar)
• "Top signal" callout with the most recent dated signal across types
• Expandable detail section (jobs/news/financing/tech/similar lists)
• Filter pills (All / vertical) + sort pills (density / recency / vertical)

Quick reference

# After signals:fetch has populated the cache for the domains you care about
python3 scripts/predictleads-dashboard.py \
  --domains personio.com,oysterhr.com,...,mirakl.com \
  --leads-json /tmp/leads.json \
  --out ~/Desktop/predictleads-dashboard.html
open ~/Desktop/predictleads-dashboard.html

Common pitfalls

• Empty cards: signals haven't been fetched yet. Run signals:fetch --domain X first.
• News headlines blank: PredictLeads news payloads use summary not title. The template's display logic falls through payload.title || payload.headline || payload.summary.
• Tech stack shows blanks: technology names live in JSON:API relationships.technology.data.id resolved via included[]. The normalizer in predictleads-enrichment.ts already promotes payload.technology to a top-level string. Older signals fetched before the normalizer fix may have empty tech rows; re-fetch with --no-cache.
• Similar companies show only score: same root cause — re-fetch with --no-cache to populate the similar_company field with the resolved domain.

Required env

None for generation (it's local-only). The signals must already be cached, which means PREDICTLEADS_API_KEY + PREDICTLEADS_API_TOKEN had to be set when the cache was populated.