← Back to Portfolio

Engineering Case Studies

Five Systems, Walked Through End to End

Most of what I have shipped professionally lives behind a login or inside a private repository, so a list of bullet points is usually where the conversation stops. These case studies are the alternative: a full walk through the systems I designed and built at Strollr and CollabWORK, the QuickFix side project, an earlier four-surface restaurant platform, and the trading systems I build and run — the architecture, the data flow, the AI systems, the failure modes I had to design around, and the reasoning behind each decision. Each one is split into sections you can step through in order, with the system diagrams that make the shape of the thing obvious, and expandable deep dives wherever the interesting detail lives below the surface.

Strollr

Founding Engineer · May 2024 – Present

Two consumer products, eight deployable services, one engineer

Strollr helps parents in Australia find nearby activities for children aged nought to five; Lil Strollr is the AI companion that sits alongside it. I built both products end to end — the React front ends, the layered TypeScript APIs, the semantic venue search, the conversational agent deployed on Vertex AI Agent Engine, the subscription metering, the admin tooling, and the Cloud Build pipelines that ship all of it. The repositories are private, so this case study walks through the architecture and the engineering decisions rather than linking to source.

TypeScriptReactNode / ExpressFirestoreVector SearchGoogle ADKVertex AIClaude SonnetStripeCloud Run

CollabWORK

Software Engineer · November 2024 – August 2025

An AI job-data pipeline holding five systems consistent

CollabWORK is an AI-powered job search platform serving the US, Canada and Mexico. I built the ingestion and enrichment backend behind it: a fully asynchronous pipeline that streams large XML job feeds, deduplicates them with fuzzy matching and union-find, enriches every listing through Gemini across seven Google Cloud regions in parallel, routes results by model confidence through a human review queue, and then reconciles five separate downstream systems so the live app, the search index and the analytics warehouse never disagree. That backend is public, so this case study links directly to the source.

PythonFastAPIasyncioPostgreSQLGemini 2.5Vertex AIBigQueryUpstashRapidFuzzDocker

QuickFix

Side project · public repository

Three roles, one workflow, enforced at the database

A rental maintenance app coordinating customers, property managers and tradespeople through an eight-state job lifecycle where each transition belongs to exactly one role. The client talks to Firebase directly, so authorization lives entirely in the security rules rather than the interface — including the rule that lets an agent claim a job from their own trade pool without letting them assign it to anyone else. Ships to web and Android from one codebase, and was deliberately hardened for public release from a private original.

Next.js 15TypeScriptFirebaseFirestore RulesCapacitorAndroidRadix UIGoogle MapsFCM

Restaurant Automation Platform

Earlier work · architecture only

Four applications, four different users, one API

A multi-tenant restaurant operations platform built across four front ends over a shared Node and MongoDB API: an installable customer PWA for discovery, reservation and ordering; a waiter and kitchen terminal for table service; a restaurant back-office covering inventory, purchasing, costing and profit and loss across sixty-seven custom screens; and a platform operator console for venue onboarding, vendors and support. It is also the oldest project here, and its case study ends with a retrospective naming exactly what its authorization model gets wrong and how that shaped everything built afterwards.

ReactNode / ExpressMongoDBMongoosePWAWorkboxStripeMaterial UIMulti-tenant

Algorithmic Trading Platform

Live · architecture only

Three asset classes, real orders, and a research record mostly made of rejections

Three independent trading systems I build and run — NIFTY index options, crypto perpetuals across four symbols, and Indian equity momentum — placing real orders since August 2026. Behind them sit research labs whose job is to destroy a signal before it reaches live code: 239 research commits, numbered experiments, walk-forward validation across three years of data, and a record where the headline results are mostly decisive negatives. The case study covers the safety machinery that live capital demands, real production incidents including a filled position that went untracked for a day, and the research record in full.

PythonasyncioSQLite / WALFastAPIsystemdWalk-forwardLive executionTailscaleTelegram