Five legacy systems, one platform, 40% less infrastructure

Context

OptoCeutics builds a digital health and neurotech platform serving 50K+ users out of Copenhagen. I joined to lead platform engineering, and inherited five legacy systems with overlapping data models and a Lambda estate that had grown one function at a time.

Nothing was broken exactly. It was just that every change touched more places than it should, and nobody could hold the whole thing in their head.

The problem

Fragmentation has a cost that does not show up as an incident. It shows up as velocity.

• Five systems, five versions of the truth. The same clinical and user concepts were modelled differently in each, so every integration was a translation layer.
• A Lambda estate nobody owned end to end. Functions accumulated faster than they were retired. Operational surface area grew with every feature.
• Infrastructure cost tracking the sprawl rather than tracking usage.
• No analytics layer. Leadership questions about the product took an engineer and an afternoon, every time.

Constraints

• Digital health, so data handling requirements are not negotiable and the tolerance for production incidents is effectively zero.
• The team had to keep shipping product through the consolidation. No freeze.
• Ten people across engineering, design, and QA, which is enough to do this properly and not enough to do it twice.

Approach

Consolidate the data models first. Everything else was downstream of the five systems disagreeing. Unifying the models across them produced a 40% infrastructure cost reduction, roughly $8K per month, mostly by eliminating duplicated storage and the translation machinery between systems.

Then reduce the estate. Re-architecting the fragmented Lambda platform cut operational surface area by 65%. The measure that mattered was not function count but how many places a single change had to touch.

Build the real-time pipeline properly. A streaming ETL pipeline on Apache Kafka and AWS Lambda, persisting to Parquet in S3, holding ingestion latency under 200ms. Parquet on S3 as the landing format meant the analytics layer could be built on top without another migration later.

Make the data answerable. An analytics stack on BigQuery and Looker Studio, delivering 12 executive dashboards. The point was not the dashboards. It was that questions about the product stopped consuming engineering time.

Run the pod deliberately. Six engineers, two designers, two QA. The process changes that moved the numbers were unglamorous: smaller PRs, clearer ownership boundaries, and a review expectation that made stale branches visible.

Outcome

Feature work continued throughout across React, Django, NestJS, and Flutter on AWS.

What I took from it

PR turnaround was the highest-leverage number. Going from three days to under four hours changed more about how the team worked than any architectural decision I made. Long review cycles create large branches, large branches create risky merges, risky merges create the incidents. Fixing the cycle time fixed things downstream that looked unrelated.

Consolidation is a cost argument before it is an engineering one. Framing the work as $8K per month made it straightforward to get the time for it. The same proposal framed as reducing technical debt would have stayed in the backlog.

Zero incidents across a year of consolidation was a team property, not a personal one. Clear ownership and small changes did that, not caution.