Strollr

Two Consumer Products, Eight Services, One Engineer

Section 1 of 5

Strollr helps parents and caregivers in Australia find things to do with children aged nought to five — parks, paid classes, free community events, at-home activities — near wherever they happen to be standing. Lil Strollr is the second product built on the same foundations: a conversational assistant, Lil Joey, that answers the questions a tired parent actually asks, remembers the family it is talking to across sessions, and ships as both a web app and an Android app from a single codebase.

I joined as the founding engineer in May 2024 and have built both products end to end since: the React front ends, the layered TypeScript APIs behind them, the semantic venue search, the agent deployed on Vertex AI Agent Engine, the subscription metering that gates it, the internal admin console, the crawling and enrichment service that keeps the venue catalogue populated, and the Cloud Build pipelines that ship every one of those services. There was no platform team to hand any of it to.

The repositories are private — they are a working company's product code, not a portfolio piece — so what follows is the architecture, the design decisions, and the reasoning behind them rather than a link to source. Both products are live and can be opened below. The four sections after this one go through the frontend, the backend, the AI systems, and the infrastructure in the order you would encounter them if you traced a single request through the stack.

2

consumer products in production

8

independently deployable services

9.5K+

lines of TypeScript in the core API

105

modules in the web client

3

platforms shipped — web, Android, admin

System Architecture

Two client applications, two product APIs, and an internal admin plane, all sitting on one shared Google Cloud data and AI layer. Everything below the clients is containerized and deployed to Cloud Run in australia-southeast1, close to the users it serves.

Strollr Web AppReact + Vite + TailwindLil StrollrWeb + Android (Capacitor)Admin ConsoleReact + MUIStrollr APIExpress + TypeScriptLil Strollr APIExpress + Stripe + PostHogContent ServiceFastAPI crawlerShared Data LayerFirestore — vector-indexed venues, users, chats, subscriptions · Cloud StorageVertex AIAgent Engine (Lil Joey) · Memory Bank · Cloud Run embedding serviceDockerized · built by Cloud Build · deployed to Cloud Run (australia-southeast1)

The Eight Services

Each of these is a separate repository with its own Dockerfile, its own Cloud Build configuration, and its own deployment. Expand any one for what it does and what it is built from.

What This Project Actually Demanded

Owning the full stack, not just contributing across it

Founding engineer on a consumer product means there is no boundary to hand work across. The same week can contain a Tailwind layout bug on a venue card, a Firestore composite index that needs redesigning, an agent prompt that is producing the wrong tone, and a Cloud Build step that is failing on a stale cached layer. The architecture in this case study is shaped by that constraint — every abstraction in it exists to keep a single person able to move safely between eight services.

Building a second product on the first one's foundations

Lil Strollr is not a rewrite. It reuses the Strollr web components, the API's layered architecture and repository base class, the authentication stack, and the deployment pipeline, then adds what makes it a distinct product — the agent, Stripe, analytics, and an Android build. Getting that reuse without the two products becoming coupled was a deliberate design exercise rather than a happy accident.

Treating AI as a system, not a feature

The interesting problems in Lil Joey were never the prompt. They were: how does the assistant remember a family across sessions without re-sending a transcript every turn; how do you keep the model swappable when the provider is a moving target; how do you see what the model actually did in production; and how do you meter access to something whose marginal cost is real. Each of those got a structural answer rather than a workaround.

Designing search that respects physical distance

Semantic search on its own returns the most conceptually similar venue, which is useless if it is four hundred kilometres away. Semantic search discarded in favour of a radius filter returns whatever happens to be nearby, which is useless if it is not what was asked for. The venue search resolves that as two ordered stages — recall by meaning, then rank and cut by real geographic distance — which is covered in full in the AI Systems section.