DEMO MODE— Read-only view with masked personal data

Pipeline

Phases, live stats, run history, and reference documentation.

Next run

in 12h 58m

COMPLETED

Pipeline Phases

Runs daily at 09:00 Europe/Bratislava. Each phase runs in sequence.

Run History

No pipeline runs recorded yet.

Reference Documentation

Normalizer Rules (Phase 1)

Rule-based modules that detect and fix contact data issues. No AI involved.

AI System Prompt (Phase 2)

This is the system instruction sent to Claude Haiku for reviewing medium-confidence changes.

You are a Google Contacts cleanup assistant. You analyze contacts and suggest corrections. You specialize in Slovak and Czech names, diacritics, phone numbers, and duplicate detection. RULES: - ALWAYS respond in JSON format - Slovak/Czech names always with diacritics - When uncertain, prefer keeping the original - Confidence: 0.95+ certain, 0.70-0.90 probable, below 0.60 speculative

Model: claude-haiku-4-5-20251001 | Cost cap: $3/day | Also loads instructions.md (editable rules) and memory.json (learned patterns)

Learning System (Memory)

The system learns from your Review decisions to improve future confidence scores.

Learning sourceYour approve/reject/edit decisions in Review

MethodBayesian smoothing — blends base confidence with your approval rate

Formulaadjusted = (base × 10 + approval_rate × decisions) / (10 + decisions)

Min decisions5 before adjustment kicks in

Confidence rangeClamped to 0.30 — 0.98 (never fully certain or fully rejected)

Categories26 rule types tracked independently (diacritics, phones, tobedeleted, etc.)

FollowUp Scoring Formula (Phase 4)

How contacts are ranked for reconnection. Three additive components.

Interaction Score

How much you've interacted and how long ago.

interaction_count x months_gap

count: 0 (none), 1 (email OR meeting), 2 (both)

LinkedIn Score

Signal from their LinkedIn profile.

job_change30

active10

profile3

no_activity0

Completeness Score

How much info you have on the contact.

signals x 2.0

Signals (0-4): email, phone, org, LinkedIn URL. Max: 8 points.

Confidence Thresholds

90%+

HIGH

Auto-applied

60-89%

MEDIUM

Goes to Review

<60%

LOW

Dropped silently

Safety Controls

Emergency Stop — Config page pause button halts the pipeline immediately.

Cost Cap — AI spending capped at $3/day. Pipeline continues without AI if limit reached.

Change Limit — Max 200 auto-applied changes per run.

Backup — Full contact backup created before every Phase 1 run.

Nothing is permanent — Deletion candidates are only flagged for review, never auto-deleted.