Skip to lesson
Exit
The Interview Loop1 / 2

2 min lesson

The practical project (decision round)

Work through this case: "You have the paid practical project: build a GTM enrichment-to-routing workflow. Name the four things that most separate a production-grade submission from a demo toy." Say what you would do and why.

Step 1 of 2

The decisive block: build or design a real GTM system, often as a paid hands-on project with the team. Treat it as production, not a demo toy, because the reviewers are the people you'd ship next to.

Shape A - orchestration workflow

A Clay (or Zapier/Unify/n8n) table that ingests leads.

Waterfall enrichment → ICP/intent scoring → routing.

Dedup, a score threshold, owner assignment with fallback.

A measurable output: enriched + routed records, match rate.

Shape B - AI agent / automation

An agent that acts against APIs or an MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. server.

LLM classification or personalization where it earns its place.

RAG to ground the agent in real account/product context.

Idempotent actions, retries and a clear audit of what it did.

  1. 1Scope to one demoable slice. One trigger, one enrich-and-score path, one routed output that runs end to end - working and explained beats grand and half-built.
  2. 2Handle dirty data first. Normalize inputs, dedup and route malformed records aside instead of crashing; reviewers look for this before any clever feature.
  3. 3Make it reliable. Respect provider rate limits, make writes idempotent so a re-run doesn't double-route and retry transient failures.
  4. 4Add AI where it earns its place. LLM classification, personalization or an action-taking agent - used because it removes real toil, not for show.
  5. 5Document and measure. A short spec/runbook plus a dashboard or metric (match rate, % routed correctly, hours saved) proves the self-sufficient-team value.
AI column / agent step: classify and personalize, but keep it grounded and idempotent
// Clay-style AI column or agent step: qualify + draft, with a guardrail
{
  "input": { "company": "{{company}}", "signal": "{{page_viewed}}", "firmographics": "{{enriched}}" },
  "task": "Classify ICP fit (A/B/C) from firmographics + signal. If A or B, draft one personalized opener grounded ONLY in the provided context. If C, return action=nurture and no draft.",
  "rules": [
    "Never invent facts not present in input (no hallucinated headcount/funding).",
    "Return strict JSON: { tier, action, reason, draft? } so downstream routing can branch.",
    "Same input must yield the same routing decision - idempotent, cache by company id."
  ]
}
Learn more

Full explanation

The reviewer's checklist

The reviewer's checklist

They skim for: does it run end to end, does it survive dirty/duplicate data, does it respect rate limits and re-run safely, is the AI used where it earns its place rather than as decoration and is there a runbook plus one metric that proves it works. Optimize for those before any extra feature.