// ai-stack · spec

Every pipeline below is real and runs in production.

Documentation-first: components, stages, latency, cost, and guardrails. If a number is missing, it means we don’t have it cached yet — ask and we’ll pull it.

PIP-01

Receipt Pipeline

Gmail → Vision → n8n → QuickBooks

Expense receipts auto-label, OCR-classify, and write back to QuickBooks with a human review pass.

Stage	Component	Notes
Stage 1	Gmail auto-label driver	5s runtime · 60 labels/run · 0 failures after week 2
Stage 2	Vision OCR + classification	OpenAI Vision · vendor + category classification
Stage 3	QuickBooks write-back	Webhook-driven · human review queue before commit

$throughput: 60 emails in 5s
$failure rate: 0% after stabilization

PIP-02

Lead Triage

n8n → AI scoring → CRM → Telegram

Every inbound email is scored, labeled, written to the CRM, and alerted to the assigned owner.

Stage	Component	Notes
Stage 1	Inbound capture	n8n webhook · multi-channel email
Stage 2	AI scoring	priority + category model inference
Stage 3	CRM + Telegram	field updated · alert to assigned owner

$latency p95: 90s end-to-end
$cost per lead: ~$0.003 (mostly model inference)

PIP-03

Deal Flow Routing

inbox → AI agent → alert

Deal-relevant inbox traffic is reviewed by an AI agent, then routed to a human with full context.

Stage	Component	Notes
Stage 1	n8n webhook	inbox trigger → agent review queue
Stage 2	Hermes review	AI agent triages and drafts response context
Stage 3	Telegram alert	chatId 8555469436 · human-in-the-loop

$state: production
$guardrail: agent NEVER writes to CRM directly

// guardrails

!>Every pipeline ends in a human review or alert. No agent writes to the CRM directly.
!>Failures are counted, not hidden — the metrics table on the landing page is a cached readout of real counters.
!>Secrets and credentials never live in workflow code; they’re injected at runtime.

Read the spec. Browse the stack. Then, if you want to ship:

[ Send scope ]

reply by next business day