Nine ventures, each with its own SDKs and event habits, produce data that no shared system can trust: legacy events arrive without IDs or schema versions, naming drifts between teams, and anything built on top — profiles, segments, campaigns — inherits the mess. And once the data is trustworthy, acting on it still needs guardrails: marketing reach across millions of users has to come with approvals, frequency caps, and quiet hours.
Ripple
Tapsi’s customer data platform: one canonical event pipeline across nine ventures, feeding customer profiles, analytics, and governed campaign delivery.
- Role
- Platform Engineer
- Year
- 2026
- Stack
- Go · Kotlin · Kafka · PostgreSQL · gRPC · Kubernetes
- Team
- Tapsi Data Platform

Ripple is Tapsi’s customer data platform. Behavioral events from nine products — SuperApp, Cab, Box, Market, Food, Shop, Doctor, Garage, and Bimeh — flow through one canonical pipeline into customer profiles, analytics, and audience segments. I build its backend services, from the ingestion gateway to campaign delivery.
Problem
Approach
- Canonical event catalog: Versioned event definitions shared by every venture SDK — one envelope shape from client to storage.
- Ingestion gateway: A Go edge service authenticates ventures, validates every payload against the catalog, and routes to Kafka deterministically.
- Stream enrichment: A Kotlin Kafka Streams processor wraps events in the canonical envelope and attaches session and context data.
- Identity resolution: Anonymous and known activity are stitched into customer profiles that feed analytics and segmentation.
- Governed campaigns: Segments are approved at an exact revision, frequency caps are enforced per venture, do-not-disturb schedules are honored, and every action is auditable.
The part that matters most: data integrity
Everything above the pipeline is only as good as the events underneath. Ripple’s most important property is that nothing uncanonical reaches storage: legacy v1 traffic is normalized at the ingestion edge — event IDs backfilled, schema versions stamped, field names canonicalized — so a silent class of data loss is eliminated without forcing a single legacy SDK to migrate. Downstream systems can finally trust their input.
Status
In active development since June 2026 — five months in and ongoing. Recent work spans campaign execution and governance, and the pipeline itself keeps growing with the ventures on it.
