Quality engineering · Test automation · Release decisions

Samuel Pangestu
G Tobing

I help software teams decide what to test, automate the checks that repeat, and turn test results into release decisions, including when one user journey spans several teams.

Lead Quality Engineer at PERURI (INA Digital) · leading a team of 30 QA engineers · in QA since 2013 · previously Stockbit / Bibit and Megaxus

Cross-team regression · Release coordination

One release check for journeys that cross several applications

When each team tests only its own application, the handoffs between applications can go unchecked. I coordinate shared regression checks so a release is judged on the whole journey.

Context
A superapp integration that consolidates 14 standalone applications into one platform, plus 3 new application projects, each built by a squad of 10–20 developers.
Problem
A user journey passes through applications owned by different teams. Each part can pass its own tests while the journey still fails.
My role
Lead a team of 30 QA engineers, coordinate shared regression checks across squads, and run one release review.
Approach
Map each shared journey end to end, include the handoffs between applications in the regression scope, and use contract testing and service virtualization so squads can test in parallel.
Release review
Four questions: scope, test evidence, open risks, and the decision with its reason. The template is below.
Decisions and trade-offs
Two calls shaped the work. First, judge a release on the whole journey, not per application — it costs more coordination across squads, but testing each application alone leaves the handoffs unchecked. Second, let squads test in parallel with contract testing and service virtualization instead of waiting on dependent services — faster, but a virtualised service can drift from the real one, so the contract is the check that keeps it honest.
How I verified it
Shared regression checks run before the release review, and the review itself asks for scope, test evidence, open risks, and the decision with its reason. Once squads could test in parallel, the cross-team integration test cycle went from about two weeks to about three days.
Example
A journey can pass inside one application and still fail at the handoff — for instance, a field the next application expects is missing or formatted differently. The shared checks catch that before the release review.
Outcome
Post-release incidents down 50% through pre-release validation of integrated journeys.

One journey, several owners

  1. Application ATeam A
  2. Application BTeam B
  3. Application CTeam C

Handoff checks between applications

One release reviewScope · evidence · risks · decision

Test automation · Shared foundation

One automation foundation every project builds on

When each project builds its own test framework, the same problems get solved many times and results look different everywhere. I built two shared core libraries and the pipeline around them, so every project writes, runs, and reports tests the same way.

Context
Several test projects at PERURI (INA Digital), each with its own framework, its own reporting, and the same setup problems solved again each time.
My role
Built the two shared core libraries and the pipeline around them, so projects share one way to write, run, and report tests.
Decisions and trade-offs
First, one shared core instead of a framework per project — it costs upfront build time and every project upgrades together, but it stops the same setup problems being solved many times. Second, release the cores as versioned packages from main, so a project can stay on a version that works and upgrade deliberately instead of drifting on its own copy.
How I verified it
Tests run in the pipeline on every deployment through Jenkins quality gates, with Allure reports and notifications to the team. Coverage and execution are tracked in the QA Test Management System, and automated tests cover 99% of critical user journeys.
  1. Versioned package

    Build on a shared core

    API Test Core (Java 17, RestAssured, Cucumber) and Web Test Core (TypeScript, Playwright, Cucumber) provide authentication, request and assertion steps, schema validation, multi-environment config, and test data from Google Sheets. Each merge to main publishes a new version.

  2. Running

    Run in the pipeline

    A Jenkins shared library runs the tests, stores the Allure report, and notifies the team in Google Chat.

  3. Ready for review

    Review the results

    Teams see which checks passed or failed and track coverage and execution in a QA Test Management System built on Google Sheets.

Outcome: 99% of critical user journeys automated with one framework for web, mobile, and API.

AI-assisted testing · Human review

AI Agent Speed: test triage with a person in the loop

When an automated test fails, someone has to work out whether the product, the test, the test data, or the environment broke. At PERURI (INA Digital) I'm building AI Agent Speed, an agent workflow that gathers evidence first and stops for a reviewer before changing what counts as a pass.

My role
Designing and building AI Agent Speed: the agent workflow, its evidence-first steps, and the review gates that stop it changing what counts as a pass.
Decisions and trade-offs
First, the agent gathers evidence before it guesses, and it may only restore test data, rerun, or propose a repair — it never changes what counts as a pass on its own. That costs reviewer time on every borderline case, but it keeps a wrong repair from silently passing a broken product. Second, verify locally before claiming anything live: four local suites with controlled inputs, and live accuracy left unmeasured rather than guessed.
How I verified it
See “What has been verified so far” below: synthetic replay (illustrative only), 44/44 local workflow checks, and live model evaluation not measured yet.
  1. 01

    Collect evidence

    Read the trace, logs, and request ID before guessing. An API 500 sends it to the server logs for that request.

  2. 02

    Classify the failure

    Decide whether the product, the test, the test data, or the environment is the likely cause.

  3. 03

    Act within limits

    Restore test data and rerun, or propose a repair such as updating how the test finds a renamed button. The original checks stay in place.

  4. 04

    Stop for review

    A change to what counts as a passing result waits for a reviewer. Without enough evidence, the agent escalates instead of assigning a cause.

Agent decision map · synthetic examples

  1. Failed testtrace · logs · request ID
  2. Collect evidence
  3. Classify the failure
  • Test data missing

    Restore setup, rerun

  • Button renamed

    Propose locator fix

    Reviewer approves

  • Pass rule would change

    Reviewer decides

  • Timeout, no trace

    Escalate with evidence

Agent may act Person decides

What has been verified so far

  1. Synthetic replay

    The agent run against made-up failures like the ones in the map. Shows behaviour, not accuracy.

    Illustrative
  2. Local workflow tests

    Four suites: evidence handling, pipeline verification, approval boundaries, and the eval runner. Controlled inputs.

    44/44 passed
  3. Live model evaluation

    Accuracy on real failures from real pipelines.

    Not measured yet

AI-assisted test generation, already in use

API Test Core turns PRD documents into Gherkin scenarios through the Gemini API, next to Postman-to-Gherkin and K6 script generators. Test creation time down 80%.

Experience

From game QA at Megaxus Infotech through mobile, API, and performance testing at a consumer investing app to leading 30 QA engineers across applications at PERURI (INA Digital).

  1. Game QAMegaxus
  2. QA AnalystMegaxus
  3. Test EngineerStockbit / Bibit
  4. Senior QEPERURI
  5. Lead QEPERURI
Feb 2013Now · 13 yr 8 mo in QA
  1. Jan 2025 – present

    Lead Quality Engineer · PERURI (INA Digital)

    Lead 30 QA engineers on a superapp integration of 14 applications plus 3 new projects. Built shared automation core libraries, introduced AI-assisted test generation, and cut post-release incidents by 50%.

  2. Oct 2024 – Jan 2025

    Senior Quality Engineer · PERURI (INA Digital)

    Led 15 QA engineers. Set QA standards and Jenkins quality gates that cut release cycle time by 50%, and introduced K6 performance testing.

  3. May 2022 – Oct 2024

    Test Engineer · Stockbit / Bibit

    Built mobile automation with Appium and Cucumber and web automation with Katalon, cutting testing time by 70%. Ran K6 performance tests.

  4. Feb 2013 – Apr 2022

    Game Quality Assurance / Tester, then Quality Assurance Analyst · Megaxus Infotech

    Planned test cases and coverage from requirements and executed them on live and test servers; investigated and documented game bugs and exploits, collected user feedback for improvements, and managed the game community around events and updates. Assisted with automation for selected modules.

Products · Built and run by samuelpgt

Products

Software I design, build, and operate on my own, outside my QA role. Every product shares one design system for UI and writing.

HariBahagia

Live · Wedding planning · Bahasa Indonesia

A wedding planning web app for couples, their families, and wedding organisers. One shared project holds the checklist and calendar, budget and payments, vendors, event rundown, and guest list.

  • Sharing by role: partner, view-only family, WO admin, and WO staff. The budget stays hidden from roles that should not see it.
  • Vendors get read-only brief links without an account. Prices and internal notes never appear in a brief.
  • React and Supabase (Auth with row-level security), served from a Cloudflare Worker.
Open HariBahagia

Design system · Shared by every product

One design system for every product I build

sp.css holds the tokens, components, motion, accessibility rules, and writing guide that HariBahagia is built on. A new product links one stylesheet, picks a style and an accent, and goes straight to features.

Joy

Personal moments, like HariBahagia. Soft dark, serif headings, rounded shapes, doodles.

Focus

Work tools, like an accounting system, and this site. Deep black, sans, sharp corners, no decoration.

Open the design system

Initiatives · Tools · Templates

Key initiatives

At a glance

AI Agent Speed
AI-assisted test triage with human review gates. PERURI (INA Digital). Outcome: failures reach reviewers pre-classified with evidence, instead of as raw logs — and the agent never changes what counts as a pass on its own.
API Test Core and Web Test Core
Shared automation libraries for API and web tests, released as versioned packages. Private. Outcome: two versioned packages shared across test projects, so API and web suites reuse the same auth, assertions, and test-data plumbing.
AI-assisted test generation
PRD-to-Gherkin with the Gemini API, plus Postman-to-Gherkin and K6 script generators. Outcome: PRDs become reviewable Gherkin scenarios with 80% less test creation time.
Performance Testing Calculator
Plans load tests with Little's Law: single-API and end-to-end modes, test levels from 70% to 150%, and k6 instance sizing. Outcome: a working planner, live on samuelpgt.com, that sizes virtual users and k6 instances from a target request rate.
K6 Performance Test Framework
Smoke, load, stress, soak, and breakpoint scenarios with HTML reports. Outcome: one framework covering smoke through breakpoint testing, with shared config and HTML reports for every run.
Jenkins QA Library
Pipeline utilities for test runs, Allure reports, report storage, and notifications. Outcome: pipeline setup shared as one reusable library — test runs, Allure reports, object storage, and Google Chat notifications.
QA Test Management System
Test cases, execution tracking, dashboards, and team KPIs on Google Sheets and Apps Script. Outcome: cases, execution, dashboards, and team KPIs in one Sheets-based suite, deployed by CI.
QA Document Templates
Eight templates from test plan to test closure report. Outcome: the full test-documentation lifecycle covered by eight templates, plus a published go/no-go checklist.
QA Platform Local Recorder
Chrome extension that records browser actions into Playwright drafts for the QA Platform, with entered values redacted before upload. Published on the Chrome Web Store (v0.3.6). Privacy policy. Outcome: published on the Chrome Web Store (v0.3.6) — browser actions become reviewable Playwright drafts, with entered values redacted before upload.
Try it: Load Test Planner120 req/s → 20 virtual users

A quick version of the Performance Testing Calculator. Estimate how many virtual users a load generator might need to start a test at your target request rate.

Starting estimate

20 virtual users

120 requests/s ÷ 3 requests per iteration = 40 iterations/s. At 0.5 seconds per iteration, about 20 virtual users are needed concurrently.

This assumes a steady arrival rate and a fixed number of requests per iteration. Use the duration of the entire iteration, including any work or delays inside it. Validate the estimate with a trial run and allow extra capacity for slower iterations. Read the k6 allocation guide.

Release Readiness Checklist

A template for recording what was tested, what risks remain, and why a release should proceed or wait. It complements the eight QA document templates.

Download the Markdown checklist