{
  "video": {
    "id": "VjEP0xqTUI0",
    "title": "The Building Blocks of GTM Orchestration — Arman Vaziri, Ramp",
    "duration": 1195,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "VjEP0xqTUI0",
    "title": "The Building Blocks of GTM Orchestration — Arman Vaziri, Ramp",
    "one_liner": "Ramp's product/growth engineering lead walks through the concrete stack behind \"go-to-market orchestration\" — an internal CDP on Postgres+Kafka fed by dbt/Snowflake reverse ETL, Temporal-based durable agent threads, a Turbopuffer vector layer over unstructured sales data, and a user-editable skill library — arguing you get there by solving one team's vertical problem at a time, not by designing the perfect system up front.",
    "summary": "Arman Vaziri argues the bottleneck in go-to-market isn't ideas — it's everything after the idea: pulling an audience, generating the artifacts, and convincing teams to adopt them. The goal is to describe an intent (\"offer Pro V1 golf balls to golfers at East Coast construction companies\") and have it fan out automatically into audiences, outbound sequences, ad and web creative, and in-app notifications. He shows the building blocks Ramp actually built — an internal customer data platform, durable agent execution on Temporal, embedded unstructured sales data with hybrid search, and a customizable skill library — using nightly pre-meeting briefs for account managers as the worked example. His thesis is that these vertical, single-team builds are the foundation for multi-team, multi-channel orchestration; you can't skip to the orchestration layer.",
    "key_points": [
      "The problem framing: good ideas are abundant across product, data, engineering and go-to-market; the cost is coordination and distribution, which otherwise moves \"on the order of months.\" Reps are in back-to-back meetings and can't absorb operational burden.",
      "They built an internal CDP: CRM, product, enrichment, web and interaction data (emails, meetings, calls, page views), internal propensity models (e.g. likelihood to attach procurement or treasury) and external signals like funding announcements. Real-time events land on a Kafka topic; Postgres backs it for transactional guarantees, referential integrity across CRM/product/third-party entities, and provenance metadata. Offline compute runs in dbt/Snowflake and comes back via reverse ETL into the same layer.",
      "Because Ramp's addressable market is effectively the entire US (now expanding internationally), online batch jobs pre-compute, pre-process and pre-ingest enrichment data via API calls for both prospects and customers.",
      "Tactical rule: solve for one team first, then scale horizontally. Outbound and meeting prep are shared across teams; things like QBR generation are isolated to one.",
      "Worked example — pre-meeting briefs for AMs: pipe in meeting events, hydrate, and fuzzy-match attendee emails and meeting titles back to accounts. That mapping is a \"sneaky hard\" problem at Ramp because the same email can act on behalf of multiple businesses, so the resolution is persisted once rather than recomputed by every downstream consumer.",
      "Durable execution is built on Temporal and is agnostic to the trigger: every agent run is a durable thread, each tool call and model call is an activity, so a dead worker resumes with accumulated state instead of reprocessing the thread. It also gives config-scoped tool access per agent and human-in-the-loop pause/resume.",
      "Unstructured data (meeting transcripts, emails, notes, enablement material, product knowledge, playbooks) is chunked, embedded and stored in Turbopuffer; agents do vector + attribute + keyword search scoped to a specific account rather than pulling the full raw corpus into context.",
      "A skill library lets individual users specify their own brief format and the information they care about in text — Vaziri credits this customization for adoption. The nightly background agent fans out per account using the Postgres CDP, the vector DB, system-owned meeting-prep skills, and user custom instructions.",
      "\"GT MCP\" exposes the exact same tools and skills the background agents use to employees, so people build their own agents and automations. When someone connects, they're effectively reporting a problem and its solution — the team then productionizes their prompts, skills and vibe-coded apps and distributes them to everyone with the same problem.",
      "The orchestration end state funnels campaign intent into Ramp Revenue, their internal app: build the audience, generate personalized copy and sequences for SDRs, spin up landing pages and creative, and have the channel owners review and sign off. Agents hold multiple campaign options as a multi-armed bandit (explore vs. known returns), with guardrails for compliance rules, rules of engagement, and not repeating the same touch.",
      "On starting small (Q&A): three years ago it was two people using GPT-3.5 to put personalized copy into outbound sequences."
    ],
    "takeaways": [
      "Fix the data substrate before the agents: one consistent entity layer (they used Postgres for referential integrity plus Kafka for real-time events and reverse ETL from Snowflake/dbt) is what makes coordinated action across channels possible at all.",
      "Persist hard resolution work — like fuzzy-matching meeting attendees to accounts — once, at the data layer, so every downstream agent inherits it instead of recomputing.",
      "Run agents on a durable execution engine (Temporal) with each tool and model call as an activity, so failures resume mid-thread rather than reprocessing from the start.",
      "Don't dump the raw corpus into context: chunk and embed unstructured sales data and let agents do scoped hybrid (vector + attribute + keyword) retrieval — it's both cheaper and faster.",
      "Expose your agent tools to employees over MCP and treat their homegrown automations as a product backlog — their prompts, skills and vibe-coded apps tell you what to productionize and distribute.",
      "Starting from scratch: pick one narrow, real use case for one team, ship it, then mirror the pattern to other teams — don't spend a year architecting the perfect general system."
    ],
    "topics": [
      "agents",
      "go-to-market",
      "customer-data-platform",
      "durable-execution",
      "rag",
      "mcp",
      "data-engineering",
      "sales-automation"
    ],
    "tools": [
      "Ramp",
      "Ramp Revenue",
      "Temporal",
      "Postgres",
      "Kafka",
      "dbt",
      "Snowflake",
      "Turbopuffer",
      "MCP (their \"GT MCP\")",
      "GPT-3.5"
    ],
    "quotes": [
      {
        "text": "there's a ton of great ideas, you know, like everybody across product and data and engineering and go-to-market have like really good ideas for things that they want to do. And the bottleneck is kind of like everything after that",
        "at": "01:02",
        "url": "https://www.youtube.com/watch?v=VjEP0xqTUI0&t=62s"
      },
      {
        "text": "the way we tend to approach these problems is solve for one team first, then scale horizontally.",
        "at": "08:10",
        "url": "https://www.youtube.com/watch?v=VjEP0xqTUI0&t=490s"
      },
      {
        "text": "Everything is represented as a durable thread built around Temporal, representing each tool call and model call as an activity.",
        "at": "10:22",
        "url": "https://www.youtube.com/watch?v=VjEP0xqTUI0&t=622s"
      },
      {
        "text": "unstructured information is probably like the most valuable thing you're sitting on",
        "at": "11:20",
        "url": "https://www.youtube.com/watch?v=VjEP0xqTUI0&t=680s"
      },
      {
        "text": "you can't spend like a year going and building like some really complicated system architecture that like is perfect. So, you have to like piece together the vertical solutions and then stick them together.",
        "at": "19:26",
        "url": "https://www.youtube.com/watch?v=VjEP0xqTUI0&t=1166s"
      }
    ],
    "words": 4033
  },
  "summary_url": "/#VjEP0xqTUI0",
  "transcript": {
    "html": "/transcripts/VjEP0xqTUI0.html",
    "txt": "/transcripts/VjEP0xqTUI0.txt",
    "vtt": "/transcripts/VjEP0xqTUI0.vtt"
  }
}