Growlytics AI — Growth Intelligence Agent

Persistent growth intelligence that remembers goals, monitors GA4 and ads, and recommends safe actions backed by live evidence.

Comment

Inspiration

Growth teams make decisions across analytics dashboards, ad managers, spreadsheets, and chat history. Context disappears between sessions, recurring checks remain manual, and AI answers can sound confident even when source data is missing or stale.

Growlytics AI turns that fragmented workflow into a persistent, evidence-backed growth intelligence loop.

What it does

Growlytics connects GA4, Google Ads, and optional Meta Ads data, normalizes it into a tenant-scoped model, and lets teams ask questions through an agent backed by deterministic tools.

The Build Week experience focuses on three connected capabilities:

  1. Cross-session memory — user-provided goals, preferences, and workflow context persist across new conversations.
  2. Scheduled monitoring — recurring GA4 and Google Ads comparisons run with leases, recovery, manual-run support, and visible history.
  3. Safe recommendations — bid and budget recommendations include evidence, limitations, and guardrails without silently writing to an ad account.

How we built it

The application uses Next.js, TypeScript, React, Prisma, PostgreSQL on Supabase, NextAuth, and a Render deployment. Typed growth tools read and normalize connector data before the model reasons over it. Memory and schedules are scoped by both tenant and user. Connector tokens remain encrypted and server-side.

Production safety rules prevent fabricated fallback rows for connected tenants, keep scheduled jobs analysis-only, and require explicit approval and rollback contracts before any future ad-platform write.

What was built during OpenAI Build Week

Growlytics existed before the submission period. The meaningful Build Week extension added after July 13 includes:

  • deterministic, user-scoped remember and forget behavior;
  • recurring GA4 and Google Ads monitoring with recovery and run history;
  • analysis-only worker authorization and safety tests;
  • authenticated Meta read and bid-plan scoping;
  • concise agent responses that separate evidence from limitations;
  • expanded contract, isolation, regression, and deployment-readiness verification.

The dated repository history and README identify the relevant Build Week commits and distinguish the new work from the earlier platform.

How Codex and GPT-5.6 helped

Codex with GPT-5.6 was used as an engineering collaborator to map the existing architecture, trace agent and connector paths end to end, implement the memory and scheduler lifecycle, strengthen failure recovery, and add tests for tenant isolation and no-write guarantees.

It also helped find and repair a Meta account-scoping path that could select the wrong connection or make a failed sync look successful. The project was repeatedly checked with linting, type checking, Prisma validation, 259 automated tests, and a production build.

Human product decisions remained explicit: prioritize persistent workflow context, scheduled cross-channel comparison, and safe recommendations; store goals rather than live campaign metrics; never substitute one ad platform for another; and never fabricate connected-tenant data.

Challenges

The hardest part was not producing an AI answer—it was preserving truthfulness across authentication, tenant boundaries, connector failures, recurring execution, and missing denominators. Scheduling also required lease and recovery behavior so that retries would not create misleading duplicate outcomes.

Accomplishments

  • Persistent memory that survives new chats while remaining user- and tenant-scoped.
  • Scheduled monitoring that remains analysis-only.
  • Evidence-backed recommendations with explicit uncertainty.
  • Honest connector behavior for GA4, Google Ads, and Meta Ads.
  • A verified local release gate with 259 passing tests and a successful production build.

What's next

Next steps include judge-friendly test access, stronger operational observability for scheduled runs, and a separately approved write workflow with audit logs and rollback—while preserving the current analysis-first safety contract.

Built With

Share this project:

Updates

Submission history