Skip to content
Phil good
PT

Kytha, the AI data assistant in Slack

The agency's J.A.R.V.I.S. You ask in Slack the way you'd ask a teammate, and she reads the client's sales, ads and website to answer, with AI and with numbers always coming from the database.

Client
Agência DPR
Type
Automation & AIMicro SaaS
My role
Concept, identity and development
Status
Live, with AI · Next: strategy reading
Year
2026
Stack
Claude (API da Anthropic)Slackn8nSupabasePostgreSQLMeta Ads API

Built for DPR agency projects

The challenge

To check a campaign, someone had to open the dashboard, filter, cross it with the sales spreadsheet and post a screenshot in the chat. Every single time. And the Ads Manager number never matched what actually came in.

What I built

An assistant that lives where the team already talks. She brings together real sales, ad spend and the path from visit to purchase, and Claude writes the reading on top of that package. The AI interprets, but doesn't make things up: every number in the answer is checked against the data before it reaches Slack. Her name comes from the Antikythera mechanism, considered the first computer in history: bronze gears that calculated the movement of the stars. Kytha does the same with the agency's data.

Highlights

  • →Answers any question about sales, ads or the website funnel, right in the thread
  • →Reads campaigns with AI: separates fact from hypothesis and suggests up to 3 actions for the media team
  • →Checks every number the AI writes; if one isn't in the data, the answer doesn't go out
  • →Real profit comes before ROAS: it's the metric that decides whether the day was good
  • →Sends the morning summary on her own and alerts when tracking drops or an integration stops
Kytha already talks with AI, but every question stands on its own: she keeps no conversation memory, on purpose. That keeps her cheap, predictable and easy to fix. Next up is a deeper strategy reading, crossing weeks, creatives and pages.

How it works

A question in Slack
  1. 1@Kytha

    someone asks in the channel, in their own words

  2. 2n8n

    understands the request: what, which client, which period

  3. 3Supabase

    gathers sales, ads, website funnel and real profit

  4. 4Claude

    reads the data and writes: facts, hypotheses and suggestions

  5. 5Conferência

    a number that isn't in the data drops the reading, and the fixed table goes out

  6. 6Resposta

    replies in the same thread, and everything is logged

Without anyone asking
  1. 1Toda hora

    n8n checks that sales and ad data keep coming in

  2. 2Alerta

    tracking dropped or an integration stopped: alerts once, no spam

  3. 38h da manhã

    yesterday's summary in Slack, real profit first

Want something like this for your business?

Message me on WhatsApp
Next projectWebsite + class management system