01 / definerequirements
Start with the workflow
Map who uses the system, where work gets stuck, and what needs a person's approval. Set the scope before asking an agent to build.
I design systems.
I build with AI.
I'm Fery. I design applications and the workflows around them, using AI to help plan, implement and review. I set the constraints and check the result.
verify --close-on-evidenceverified
agentedit 3 files, run tests
agent"done, all tests pass"
gaterun the project's own checks …
gate✗ 2 failing → task stays open
agentfix the failures, try again
gate✓ all checks pass → closed on evidence
Task ledger
Agent says"done"
Project checkspassing
Attempts2
Statusclosed
Illustration of the rule my orchestrator enforces: an agent saying "done" never closes a task.
12 yrssoftware design and delivery
32M+lessons on edbot.ai
4 appsbuilt and shipped solo
8 yrsfully remote, 10+ countries
// how I work
From a workflow to working software.
AI helps with planning and implementation. I keep the design decisions explicit and check the result against the original need.
02 / designarchitecture
Make the trade-offs explicit
Define system boundaries and data flow. Record failure paths and release checks, so the plan explains how the system will run.
03 / directAI-assisted delivery
Give agents bounded tasks
Use AI to turn a reviewed plan into small changes. Give each task a clear scope and acceptance criteria, with isolated workspaces where needed.
04 / verifyreview and testing
Check the whole path
Review the changes and test the user's workflow. An agent's completion message is a claim to check; passing unit tests alone can still miss a broken first run.
// selected work
The design behind the software.
Human approval in marketing automation. Task isolation in coding-agent tools. Application architecture and release automation for edbot.ai.
production · 2026Claude on Bedrock
A marketing agent that can't publish on its own
I designed the path from staff updates to draft posts, with human approval required before publication. Built on AWS Bedrock.
update→draft→review→human gate→publish
platform2026
0 → 7
live tools on Solve Education's automation platform in one quarter.
personal project · Brodevorchestration
Rules for agent-led delivery
I designed a coding-agent orchestrator with isolated git worktrees and task closure based on passing project checks. Architecture decisions live in ADRs.
Rustgit worktreesverification gates
open source · mnemreleased ↗
Memory across agent sessions
Local-first memory built from agent transcripts. Events and read progress commit in one SQLite transaction, so interrupted capture can resume.
RustSQLiteClaude Code / Codex / pi
edtech at scale · 2018 to 2025edbot.ai ↗
Application architecture at learning-platform scale
I owned the frontend architecture of edbot.ai, which delivered 32M+ lessons across 10+ countries. I also introduced the team's first CI/CD pipeline and end-to-end tests.
ReactNext.jsCI/CD100K+ peak monthly users
// evaluation and verification
Selected research.
Accepted workshop poster2026 · non-archival
When a forecasting score rewards the wrong report
A benchmark can reward a forecaster for reporting something other than what it believes. We examine why, test the mechanism in simulation and practice tasks, and propose normalization fixed before the outcome.
Per-Outcome Baseline Normalization Makes Proper Forecast Scores ImproperFery Yundara Putera and Bramantya Farid Prakoso.Accepted for poster presentation at Agenthon 2026: Verifiable AI for Quantitative Finance, a workshop at NeurIPS 2026.
Workshop details ↗// side products
Four apps, from idea to release.
Personal products I designed and shipped solo, each with a different workflow and platform constraint.
Vacua
Mac cleaner with a Safe or Moderate label on every item. 9 scan categories, $12 once.
Qadha
A 30-day guided reset for prayer consistency. Donation-based, nothing paywalled.
Plareon
Daily basketball plans for 6 player types, with audio drills and an Elite tier.
Zokuu
Alarms that fire even in Doze mode or after a reboot. No account, no network.
// experience
Twelve years, mostly remote.
Solve Education!2018 to now
Senior Software Engineer. I lead the AI automation platform and built the team's AI coding-agent setup. Before that I owned edbot.ai's frontend and built the team's first CI/CD.
Xtremax2016 to 2018
Backend Developer. Maintained three live client projects for teams in Singapore and Indonesia. Certified in Sitefinity CRM.
Elephant Talk2015 to 2017
Software Developer. C# microservices and two AngularJS web apps, Spain and Indonesia.
Suzuki Indomobil2014 to 2015
.NET Developer. Dealer management system used in every main branch office.
// writing
Notes from the gate.
Close only on evidence: running coding agents I don't trustsoon
Ship more, break less: CI/CD pipelines that actually workMay 2026How we scaled edbot.ai to 32 million lessonsMay 2026Building AI-agentic workflows: lessons from EdTech at scaleMay 2026From .NET to Next.js: a 12-year engineering journeyMay 2026// work with me
Two ways to work together.
I'm open to a full-time remote role, and to project-based or freelance work. Both follow the same process: define, design, build with AI, verify.
full-time · remoteUTC+7
Join your team
Application and AI systems design, with responsibility for delivery. My working day overlaps with Asia, Australia and the European morning.
project · freelanceopen now
Build one thing, well
A project with a clear finish line: a new product, an AI workflow or a rebuild. You get a written plan before any code, working software at each step, and a handover your team can run.
AI workflowshuman approval
Automation that knows when to stop
Workflows that do the routine work and wait where a person must decide. Proof: a marketing workflow on AWS Bedrock that can't publish without approval.
productsidea to release
From idea to the app store
Web and mobile products shipped end to end. Proof: four apps I designed and released solo on macOS, iOS and Android.
teamsAI-assisted delivery
Coding agents your team can trust
Shared rules, required checks and delivery commands, so agent work passes review. Proof: the setup my team uses today.
Want to build something world‑class?
Tell me what you want to build, as a project or a full-time role. I reply within 24 hours with questions and a first take on scope.
Or email feryyp.work@gmail.com