Reinaldo Gunawan
From graduate analyst to CTO
Kick Avenue is one continuous role, founding CTO from 2017 to now, full-time from 2022. The other roles ran alongside it.
Indonesia's sneaker & streetwear marketplace
An authenticated resale marketplace with bids, offers, consignment and warehouse verification. I architected and led the entire technology stack, from the first commit to today's platform.
- Kard Avenue launched in April 2026 on Kick Avenue’s backend, 3 months after the instruction to build.
Buyers needed to trust that every pair is authentic, and sellers needed one place to list, price, ship and get paid, across the web app, the mobile app, and marketplaces like Shopee and TikTok Shop.
CTO and hands-on architect. I designed the system. Across my leadership roles I have hired over 50 people (across all roles) in engineering, QA, product and design. Kick Avenue’s current team is 4 engineers, 1 product, 1 QA and 1 designer. We use AI-assisted development and review. I still ship backend code every week (1,292 commits to the API monolith since Oct 2025).


One platform, many channels
- 7 people now ship what 12 did, with output and quality held by AI-assisted review and automated tests
- Prioritised the work that moved the business: in-house Shopee and TikTok instead of middleware, and checkout as its own service
- Set the review culture: an AI-assisted PR review workflow (Claude Code & Codex) shared across repos
- Run incident response: postmortems that end in codified probes and edge alerts
- ~1,700 PHPUnit test files in the monolith
- Feature work ships with tests and a written runbook
- Moving hot paths out of the monolith, starting with checkout
Lumen 5.5 → Laravel 12, then a returns engine on top
A 2017-era Lumen 5.5 monolith ran every money flow. It was hard to test and hard to scale, and it lacked a safe workflow for returns, refunds and relisting.
- Upgraded Lumen 5.5 → Laravel 12 / PHP 8.4 behind a test harness I built first
- Moved to Octane on FrankenPHP; fixed long-running-worker memory leaks
- Wrapped OpenSearch in a circuit breaker with cached fallbacks
- Built the returns workflow: QC → relist or refund, idempotent credits, a 1.5× refund ceiling, and the CMS screens for it
- Search can fail without taking the API down: OpenSearch sits behind a circuit breaker and cached fallbacks
- Returns, refunds and relisting follow one workflow, and a refund cannot exceed 1.5× the order


Direct Shopee & TikTok Shop integration
Replacing Anchanto with in-house sync means one inventory sells on Shopee and TikTok. Order volume and the share of GMV are available on request.

Built to fail safely
Designed and built both integrations end to end: API clients, signing, sync engine, order ingestion, CMS handover, and the technical reference the on-call team uses.
- Twin services, no shared state. Shopee item/model and TikTok product/SKU are mirrored but isolated, so one channel's outage never blocks the other
- Write guards and rate-limit backoff that honour Retry-After headers
- Adopt before create: existing remote listings (even Anchanto-era ones) are claimed, never duplicated
- Idempotent shipping: "already shipped" counts as success, so retries are safe

Carving checkout out of the monolith in Go
Checkout is where money moves and where bugs hurt most. I'm moving it into a small Go service with explicit offer states, so every payment outcome, late or out of order, lands in a defined state. Checkout is the first slice of a strangler-fig move: it leaves the monolith now, and the rest of the backend follows to Go one service at a time. We estimated at least a 30% infrastructure saving from lean services instead of scaling the PHP monolith’s nodes.
- Late or mismatched captures become a settlement or a refund, never a silent paid
- Stale offers expire and restock; a repeat submit reuses the open offer
- Kubernetes manifests checked with kubeconform in CI

A serverless Rust backend, from scratch
A windscreen-replacement business needed a fast, cheap-to-run catalogue and ordering API: windscreens searchable by make, model, year and position, priced by several merchants, plus a technician app. Branches had to run in the UK, Denmark, the USA and Indonesia from one place, so the data layer had to be global.
Only backend engineer. I built the new stack and migrated it off a legacy Python backend, and used the work to train up on Rust: data model, API, bulk catalogue imports (NAGS and Mygrant CSV), group-by-year availability, cursor pagination, multi-merchant pricing.

A clinic ERP and storefront, solo
An aesthetic clinic needed one system for appointments, treatments, stock, staff leave, payments and patient records, connected to SatuSehat, Indonesia's national health-record platform.
- Appointment calendar by clinic, staff and room, with commissions per assignee
- Orders with down payments via Midtrans, stock and low-stock alerts
- SatuSehat (FHIR) patient-record lookup; private S3 storage for medical photos
- HR: leave rules (cycles, caps, eligibility) and overtime; Indonesian/English UI


Hard coding tasks for training AI models
I wrote long engineering tasks used to train and evaluate models. Each one was solvable, verifiable, and graded on behaviour rather than a single expected diff.
- Tasks came with a reference solution I wrote and checked, plus tests the model could not see
- Graders punished shortcuts: reading the hidden tests, or a fix that only worked on the small cases
- One example was a Kafka-compatible broker in Rust, built in five stages
Small AI tools I built for real problems
A chat that answers "how do we…?" from the team's own Jira, Confluence and Slack, with citations. Docs are chunked and embedded, retrieved with pgvector, and answered by a swappable LLM.
- Pluggable LLM layer: switch Gemini ↔ OpenAI by config, embeddings included
- Resumable Slack ingestion, so a large workspace syncs in safe increments
- Raw documents archived to object storage (Google Drive, Huawei OBS or local) for re-indexing
- Answers stream token by token and cite their sources
Turns a full table-tennis match video into ready-to-post highlight clips. A spoken "highlight" plus a clap marks the rally, and ffmpeg cuts it.
- Two signals, one cue: a spoken keyword and a clap within 2.5 s merge into a single highlight
- Looks back 18 s to capture the whole rally before the cue
- Drops to a lower frame rate automatically when the API rejects or rate-limits a request
- Compresses sources over the 2 GB upload limit; every clip is at least 10 s and under 180 MB