LedgerStream
Event-sourced ledger processing 40M transactions/day
Lead Application Developer
Problem
The billing platform reconciled balances in nightly batches, so disputes took up to 24 hours to surface and finance had no real-time view of exposure.
Impact
Cut reconciliation latency from 24 hours to under 2 seconds and eliminated a class of double-charge incidents entirely.
Tech stack
- Kotlin
- Kafka
- PostgreSQL
- gRPC
- Kubernetes
The problem
The billing platform settled balances in a nightly batch job. Any mismatch — a failed capture, a partial refund, a retried webhook — stayed invisible until the next morning's run. Finance worked from numbers that were always a day stale, and disputes piled up in a queue no one could trust.
Approach
I modelled the ledger as an append-only event log. Every money movement becomes an immutable event; balances are a projection rebuilt from that log. Kafka carries the events, a Postgres projection serves reads, and gRPC exposes the query API to downstream services.
- Immutable events make every balance reproducible and auditable.
- Streaming projections keep the read model within seconds of the write log.
- Idempotent consumers make webhook retries and replays safe by construction.
Outcome
Reconciliation latency dropped from 24 hours to under two seconds. Because the log is the source of truth, an entire class of double-charge incidents — previously caused by races in the batch job — became impossible. Finance got a live exposure dashboard for the first time.