Skip to content

architecture.md

Latest commit

 

History

History
58 lines (44 loc) · 3.59 KB

architecture.md

File metadata and controls

58 lines (44 loc) · 3.59 KB
Raw

Architecture — gtm-ops 3-layer stack

This repo (gtm-ops) is the working implementation of the GTM AI Stack reference architecture — 28 tools across INPUT / BRAIN / OUTPUT, mapped to 10 GTM Builder flows. The conceptual reference lives in the wiki at C:/Users/mothi/llm-wiki/wiki/concepts/gtm-ai-stack.md (the tool catalog is its own concept; this repo is one implementation of it).

Layers

INPUT — every possible buying signal

  • CRM — HubSpot (system of record)
  • Enrichment — Clay (orchestrator) over Apollo, ZoomInfo, Clearbit, Proxycurl, Hunter, BuiltWith + Claygent AI
  • Intent — Common Room (unified 1P/2P/3P), 6sense, Bombora, G2 Buyer Intent
  • Visitor ID — RB2B / Koala / Clearbit Reveal
  • Call intel — Gong / Fireflies / Fathom / Granola
  • Competitive — Ahrefs / Meta Ads Library / BuiltWith
  • Content signal — Typefully analytics (replies, saves, DMs as intent)
  • Custom web — Scrapling (no-API gap-filler)

BRAIN — orchestration + knowledge

  • Orchestration — LangGraph (stateful + HITL) + cortex (WAT pattern skill modes)
  • Model routing — OpenRouter (Claude ↔ Haiku ↔ GPT ↔ Llama)
  • Knowledge — llm-wiki (durable) + Postgres + pgvector (structured + embeddings)
  • Evals — Promptfoo / Braintrust

OUTPUT — action

  • Dashboards — Metabase
  • Rep UI — Slack bot, HubSpot writebacks, Gmail drafts
  • Cold email — Smartlead (preferred), Instantly, Lemlist
  • LinkedIn — Dripify, HeyReach
  • Enterprise SEP — Salesloft, Outreach
  • Content distribution — Typefully (via Composio MCP)
  • Paid activation — LinkedIn / Meta / Google Ads APIs

Tool-to-flow map

Flow Tools it routes through
1. Meeting prep copilot HubSpot + ZoomInfo + Gong + Granola + Ahrefs + Meta Ads + llm-wiki → LangGraph + OpenRouter (Opus) → Slack
2. Slack action items Gong/Fathom → LangGraph → HubSpot tasks + Slack
3. CRM enrichment HubSpot webhook → LangGraph → Clay waterfall (Apollo → ZoomInfo → Proxycurl → Claygent) → OpenRouter (Haiku) → HubSpot writeback
4. Follow-up drafter Transcript → LangGraph + HITL → OpenRouter (Haiku+Opus) → Gmail draft
5. Cold-deal dashboard HubSpot → Postgres view (cold_deal_flags) → Metabase
6. Forecasting with CI HubSpot → Postgres → Prophet → Metabase
7. Win/loss analysis HubSpot + Gong → LangGraph cluster extractor → llm-wiki + Metabase + Typefully thread
8. Multi-channel outbound Clay/Common Room list → LangGraph → Smartlead + Dripify → reply webhook → HubSpot deal + Slack
9. Champion tracking Common Room + Clay Signals (job changes) → LangGraph → Smartlead + LinkedIn Ads
10. Content-to-pipeline Typefully analytics → LangGraph → Clay enrich → HubSpot lead + Slack

Why this shape

Tools are swappable (ZoomInfo ↔ Apollo ↔ Clearbit on the enrichment side). Signal sources multiply each year (new ones slot into INPUT without touching BRAIN). The orchestrator (LangGraph) owns state, retries, and HITL — not the tools themselves. Every integration lives in src/gtm_ops/integrations/ as a thin typed client with a single responsibility.

Gaps this scaffold doesn't close

  • Salesforce — Razorpay-scale orgs run Salesforce, not HubSpot. Swap HubSpot client for a Salesforce REST client; schema stays the same.
  • Production model serving — add LangSmith for tracing + eval observability in prod.
  • Access tiers — Dripify and ZoomInfo require partner/enterprise API access. Portfolio demo uses Apollo + Proxycurl as stand-ins.