Building a CSAT insights pipeline with Hasura, Claude, and Mailgun

Every week we collected hundreds of CSAT survey responses and did almost nothing with them. The scores went into a dashboard; the comments — where all the signal lives — went nowhere. This is how I closed that loop with a small, boring pipeline.

The shape of the problem

Raw survey rows aren’t insight. A pile of “the booking flow was confusing” comments only becomes useful once it’s clustered, counted, and put in front of the right person on a cadence they’ll actually keep.

So the pipeline has three jobs:

1. Pull the week’s responses out of the database.
2. Summarise them into recurring themes.
3. Deliver the digest somewhere it gets read.

Pulling the data

The responses live behind Hasura, so extraction is just a GraphQL query with a date filter:

query WeeklyResponses($since: timestamptz!) {
  csat_responses(where: { created_at: { _gte: $since } }) {
    score
    comment
    channel
  }
}

No ORM, no extra service — Hasura already exposes exactly the slice I need.

Summarising with Claude

The comments go to the Claude API with a prompt that asks for themes, not prose. The key move is forcing structured output so the next step doesn’t have to parse English:

const res = await anthropic.messages.create({
  model: "claude-sonnet-4-6",
  max_tokens: 1024,
  messages: [{
    role: "user",
    content: `Cluster these support comments into themes.
Return ONLY JSON: { theme: string, count: number, example: string }[].

${JSON.stringify(comments)}`,
  }],
});

const themes = JSON.parse(res.content[0].text);

Ask a model for JSON and it will usually oblige — but validate anyway. One malformed response shouldn’t take down the whole job.

Delivering it

The themes get rendered into a simple HTML email and sent through Mailgun every Monday morning. That’s the entire trick: the insight has to arrive before the week’s decisions get made, or it’s just archaeology.

What I’d change

The clustering is stateless — it forgets last week existed. The next version keeps a running set of themes so we can see a complaint trending, not just present. That’s the difference between a report and an early-warning system.