Vercel indexes its entire GTM memory on turbopuffer

Vercel's GTM engineering team built an AI lead agent that can access Gong transcripts, Slack channels, and Salesforce data. It's already saved over $2M.

$2M+

incremental revenue

10→1

SDR headcount

32x

ROI

I realized I could index Vercel's entire GTM corpus on turbopuffer with my credit card.

Drew Bredvick, GTM Engineer

Drew Bredvick leads the GTM engineering team that's refactoring Vercel's GTM motion. He and the team are building AI agents that qualify leads, coach deals, and analyze losses. In about 9 months, Vercel's GTM agents have collectively generated over $2M in incremental revenue.

Their first project was a deal coaching agent for Vercel's inbound sales development reps (SDRs). Drew and team first attempted to engineer context by stuffing data into the prompts, but that overflowed context windows. They tried a few pre-processing hacks to eliminate superfluous data but it became pretty clear that they'd need to build a retrieval layer to allow Vercel's GTM agents to search across the entire account conversation history and inject relevant context at runtime.

Drew chose turbopuffer to index Vercel's GTM corpus. Vercel's SDRs can now work leads with real-time coaching from AI agents that have access to the entire account history.

Why turbopuffer?

Drew ran cost comparisons between search providers. With the alternatives he considered, indexing the entire corpus of Gong calls, Slack messages, and emails was going to need a budget approval from finance that he wasn't certain he'd get. With turbopuffer, Drew could index everything on his corporate card.

Internal tool teams are a cost center by definition. I was really concerned about costs exploding as the data grew, but that hasn't been an issue with turbopuffer.

Drew Bredvick, GTM Engineer

Beyond cost, turbopuffer's API made it easy to start and simplified the implementation:

Hybrid search: turbopuffer does both full-text search and vector search, so the GTM engineering team didn't need to operate two search engines to get relevant search results
Namespace-per-account isolation: Each Vercel target account maps to a turbopuffer namespace, so each account has its own search index with isolation for data privacy
Cold/warm economics: GTM data is "human scale." Only a subset of accounts are actively being worked at any time. Active accounts get fast, cached retrieval. Dormant accounts fade to cold storage without accruing costs for in-memory indexing and idle capacity

Results

The lead qualification agent alone allowed Vercel to reduce inbound SDR headcount by 10x and move people into outbound and account executive roles. The remaining inbound team reviews AI-qualified leads with full context from the entire account history.

$2M+ in incremental revenue from the lead agent
32x ROI on the GTM engineering investment
Salespeople are 10x more efficient

turbopuffer in Vercel

Vercel's account knowledge base built on turbopuffer uses a straightforward architecture:

┌────────────┐
│    Gong    │─┐                      ╔═ turbopuffer ════════════════════════╗
│    API     │ │                      ║                                      ║░
└────────────┘ │                      ║  namespace: {salesforce_account_id}  ║░
               │                      ║                                      ║░
┌────────────┐ │        ┌──────────┐  ║  ┌────────┬─────────┬──────┬──────┐  ║░
│   Slack    │─┼─ .md ─▶│  chunk + │─▶║  │   id   │ content │ type │ meta │  ║░
│    API     │ │        │  embed   │  ║  ├────────┼─────────┼──────┼──────┤  ║░
└────────────┘ │        └──────────┘  ║  │ call-0 │  chunk  │ gong │  {}  │  ║░
               │                      ║  │ msg-42 │  text   │ slack│  {}  │  ║░
┌────────────┐ │                      ║  └────────┴─────────┴──────┴──────┘  ║░
│ Salesforce │─┘                      ║                                      ║░
│  CRM data  │                        ╚══════════════════════════════════════╝░
└────────────┘                         ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

    ┌─────────────┐
    │ Gong API    │
    │ Slack API   │
    │ Salesforce  │
    └──────┬──────┘
          .md
           ▼
    ┌─────────────┐
    │   chunk +   │
    │   embed     │
    └──────┬──────┘
           ▼
╔═ turbopuffer ═══════════╗
║ ns:{sf_account_id}      ║░
║ ┌────┬───────┬─────┐    ║░
║ │ id │content│ type│    ║░
║ ├────┼───────┼─────┤    ║░
║ │c-0 │ chunk │ gong│    ║░
║ │m-42│ text  │slack│    ║░
║ └────┴───────┴─────┘    ║░
╚═════════════════════════╝░
 ░░░░░░░░░░░░░░░░░░░░░░░░░░░

Markdown conversion: Gong calls, Slack channels (internal and external), and emails are converted to standardized markdown representations
Embed and index: Content is chunked, embedded, and indexed for vector + full-text search in turbopuffer
Namespace per account: Each Salesforce account ID maps to a turbopuffer namespace for isolation
Agent tool calls: AI agents get a search tool and generate their own vector and BM25 queries at runtime, directing them to the account-specific namespace as SDRs and AEs work a specific lead
Security and encryption: All customer data is encrypted at rest with AES-256

What's next

Drew's team is adding more data to their indexes and building more GTM agents:

Real-time objection handling: Surfacing relevant objection responses to reps during live calls using Zoom's live transcript API and turbopuffer queries
More data sources: Continuing to add context to the knowledge base as new sources become available
Multi-account large indexes: Creating new namespace schemas that allow salespeople, product leads, and engineers to ask thematic questions - like "What are the common objections from prospects that need CDN features?" - across multiple accounts