SynCORE — From 3 Months to 3 Days
The evolution of AI-assisted development. Same operator, same vision, 30x faster.
Executive Summary
The Challenge
A growing industrial services company processed hundreds of emails daily across vendors, customers, internal teams, and noise. Deterministic rules handled the obvious cases, but the real world is messy: ambiguous sender names, inconsistent email subjects, partial matches, forwarded chains with stripped context.
The question wasn't whether to automate email triage. It was how to handle the fuzzy data that breaks every rule-based system.
Phase 1: The Google Apps Script Build
The first system was built entirely in Google Apps Script, orchestrated by Gemini as the AI development partner. Over three months, hundreds of GAS routines were written, debugged, patched, and iterated into a production email triage engine called Athena Triage.
What It Did
Athena classified every inbound email into six operational folders:
The Key Innovation
AI wasn't layered on top as a chatbot or summary tool. AI was embedded into the workflow to solve fuzzy data issues. When deterministic rules couldn't classify an email — ambiguous sender, vague subject line, partial match — AI inference resolved the ambiguity and routed it correctly.
This was the insight: AI resolving ambiguity that rules can't handle, creating novel deterministic outcomes from probabilistic inputs.
The Result
Athena Triage worked. 95KB+ of active GAS code, multiple versions, continuous patches. It was one of the best systems ever built on that platform.
The Wall
Google Apps Script doesn't scale. The system hit every ceiling the platform has:
Gmail API quotas throttled triage during peak volume.
6-minute hard cap on script execution. Complex classification chains timed out.
No real database. PropertiesService has size limits. No queryable state.
GAS can call Google services. It can't orchestrate across platforms or local systems.
Every operation routed through Google's infrastructure. No data sovereignty. No offline capability.
The GAS version wasn't bad. It was excellent — and it had hit the ceiling of its platform.
Phase 2: The Claude Code Rebuild
The entire SynCORE architecture was redesigned and rebuilt using Claude Code as the AI development partner. Same operator. Same vision. Different tool, different methodology, radically different speed.
New Architecture
Phase 1: Cloud
- Google Apps Script
- Gmail API (rate-limited)
- PropertiesService (limited)
- Google Cloud execution
- 6-min execution cap
- No local state
Phase 2: Local
- Python + SQLite (WAL)
- n8n workflow automation
- Mac mini (dedicated hardware)
- No rate limits
- No execution caps
- Full data sovereignty
What Changed
Same email intelligence. Same classification logic. Same AI-embedded ambiguity resolution. But now running on local hardware with a real database, real orchestration, and zero cloud dependency. 100x faster execution, unlimited runtime, full data sovereignty.
The Timeline
Project Kickoff
Email triage problem identified. Gemini selected as AI development partner. Google Apps Script chosen as platform.
Athena Triage Takes Shape
Core classification engine built. 6-folder taxonomy established. Deterministic rules cover 70% of email volume.
AI Inference Layer Added
AI embedded into the workflow to handle fuzzy data — ambiguous senders, inconsistent subjects, partial matches. Classification accuracy climbs.
GAS Ceiling Hit
95KB+ of active code. Rate limits, execution caps, and cloud dependency block further scaling. System works but can't grow.
3 Days with Claude Code
Entire architecture redesigned. Python + SQLite + n8n on a Mac mini. Same intelligence, 100x faster, zero cloud dependency. Full data sovereignty.
The Insight
"We embedded AI into the workflow to solve fuzzy data issues and create awesome outcomes."
This is not a story about AI replacing humans. It is a story about AI resolving ambiguity that rules cannot handle — creating novel deterministic outcomes from probabilistic inputs.
Phase 1 proved the concept: AI inference inside a workflow engine can solve fuzzy data problems that break every rule-based system.
Phase 2 proved the methodology: when the tools and the operator mature together, what took 3 months can be rebuilt in 3 days. Not because the first build was bad — but because the second build carried every lesson from the first.
The methodology matured. The tools matured. The fusion works.
Ready to Build Your AI-Powered Workflow?
VindexAI builds operational intelligence systems that embed AI where it matters — inside the workflow, solving real problems.