Rafii Manggala, AI engineer

Freelance, remote, UTC+7

AI agents that do the work, not just write the code.

Most people use Claude Code to write functions; I deploy it as the engine: always-on agents that observe, decide, and act in production.

AI
Full-stack
Automation

27 more on the skill map

See what I buildGet in touch
My desk in Jakarta. It is there now, and the window shows that sky.
Skill map

What do you need built?

Seven kinds of work I get hired for, one island each. The dots on an island are the tools I use, its landmark is a real project, and its card links to the case study and the stack behind it. Four more islands hold the rest: the warehouse has all 6 case studies, the office block is Agent HQ, where the 18 agents I run have a room each, the flag on the stairs marks the 8 months that led there, and the post office is how to reach me.

    Pick an island.

    Click an island, or Tab in and use the arrow keys.

    0 of 21 tools found0 of 7 flags up

    • Full-Stack Web Apps

      From nothing to a product your people log into every day. I build all of it and put it live.

      995 schools and 12,495 users on the live system.

      Tools: React, .NET 9, PostgreSQL, TypeScript, Angular, Entity Framework, Node/Express, Django, MongoDB, Next.js, Firebase.

      Proof: K-12 Education SaaS. 18 features delivered into an existing production system serving 995 schools and 12,495 users, on .NET 9 and Angular.

    • AI Inside Your Product

      AI that does a real job for your users, not a chat box bolted onto the corner.

      Four data sources read into one score.

      Tools: Anthropic SDK, OpenAI, Whisper, Groq, Document parsing, Node/Express, .NET 9, Python, FastAPI.

      Proof: Health Optimisation Platform. An AI layer reasons across blood, DNA, DEXA and wearable data to surface compounding risk, then generates plans from it.

    • Automation That Runs Itself

      Your repetitive work happens on schedule, and you can see which step failed.

      30+ node pipeline, self hosted, runs unattended.

      Tools: n8n, Playwright, Claude Code, Klaviyo, launchd, Shopify CLI, Google Sheets.

      Proof: Content Automation Pipeline. A 30+ node self-hosted n8n graph: sources in, script and voiceover and render in the middle, publishing out the other end.

    • AI Video At Scale

      Social video produced and published at volume, with scheduling handled for you.

      600+ assets across 14 client accounts.

      Tools: Kling, Veo, Creatomate, Seedance, n8n, Whisper, TikTok and Instagram publishing.

      Proof: AI Video Production at Scale. 600+ AI-generated and AI-edited video and image assets across 14 client accounts, with anonymised sample campaigns.

    • Design And Prototypes

      A complicated idea becomes screens people understand, plus a clickable version to try.

      A live trip planner that draws its own FMCSA log sheets.

      Tools: Leaflet, Tailwind, Canvas, React.

      Proof: Spotter ELD. A regulated, paper-heavy compliance task compressed into one calm screen, log grid included.

    • Fixing Live Systems

      Something live is broken or scary to touch. I fix it with proof, behind a tagged rollback.

      9,800 stuck emails traced to the provider.

      Tools: Docker, MongoDB, Git tagged rollbacks, Playwright, PostgreSQL, Shopify CLI.

      Proof: K-12 Education SaaS. Production debugging with real schools on the system: decompiled shipped DLLs to prove a deploy matched what was tested, traced a 9.8K-email backlog to a silent provider rate limit.

    • Custom Shopify Builds

      A stock theme learns a feature Shopify apps can't sell you, built into the theme itself.

      An inline pattern editor, wired to an external API.

      Tools: Liquid, Shopify CLI, Klaviyo, JavaScript, REST APIs.

      Proof: Made-to-Measure Shopify Platform. Every read and write goes through an external REST API the theme doesn't control, inside a pure Liquid and vanilla-JS theme with no build pipeline: no React, no bundler, no shortcuts.

    • Pixel art, drawn live

      Four scenes in 9,446 lines of JavaScript with 217 tests and no image files. Every frame is computed from the clock.

      • Siang di Raja Ampat. Noon in a lagoon, seen from the waterline. Every 90 seconds the water darkens and a 15 metre mosasaur rises past you. 3,332 lines, 89 tests.
      • Senja di Kahyangan. A Balinese temple on floating islands just after sunset. A green dragon swims in from the distance every 60 seconds. 2,439 lines, 71 tests.
      • Hujan Neon. A cyberpunk street in the rain: neon signs that fail letter by letter, a maglev every 24 seconds, 770 raindrops. 2,740 lines, 46 tests.
      • Jakarta 02.00. Jakarta at two in the morning, from the canal: an MRT every 24 seconds and a meteor every 17. 935 lines, 11 tests.

      Play the four scenes on the lab page

    Skill mapWork

    Contact
    Recent work

    What I actually shipped.

    Six builds, stacked. Keep scrolling and each one covers the last.

    1. 01

      AI Video Production at Scale

      600+ assets across 14 client accountsView case study
    2. 02

      Content Automation Pipeline

      Self-hosted, zero manual postingView case study
    3. 03

      K-12 Education SaaS

      Live in production for real schoolsView case study
    4. 04

      Health Optimisation Platform

      4 data sources, one clinical scoreView case study
    5. 05

      Made-to-Measure Shopify Platform

      10 Klaviyo flows, every default email replacedView case study
    6. 06

      Spotter ELD

      Regulated compliance work, one calm screenView case study

    More I’ve shipped.

    A client app that runs behind a login, tools on GitHub, and a few things made purely in code. The agents have their own building, right below.

    • Released here

      Wake

      An 18-second film in three.js: a robot builds itself out of floating parts and wakes up.

      74 meshes, zero asset files

      Built with three.js, rendered offline frame by frame, with a score synthesized in code.

      Watch Wake with sound (opens in a new tab)
    • Live

      Employee wellbeing app

      Employees get personal growth tools. Managers get live psychosocial-risk dashboards and compliance reports.

      Aligned to ISO 45003 and Australian WHS law

      Built with React Native and Expo, an Express API, and Supabase Postgres.

      Client work in Australia. It runs behind a login, so there is no public link or screenshot.

    • Open source

      Paper Collage

      A Claude Code plugin that turns a short script into a DIY paper-collage video. The loop here is its own template.

      v1.0.0, passes claude plugin validate

      Built on HyperFrames, which renders HTML to video. MIT code, OFL fonts, CC0 photos.

      Read the code on GitHub (opens in a new tab)
    • Open source

      Trading Command Center

      Twelve paper-trading bots in one dashboard. Two models have to agree before any trade, and volatile markets are skipped.

      12 bots, 42 instruments, paper money only

      Built with FastAPI, PostgreSQL and Alpine.js, with Claude and Groq models voting on every trade.

      Read the code on GitHub (opens in a new tab)
    • Senja di Kahyangan in pixel art: temple islands floating above the clouds at dusk, with a dragon kite overhead.In the lab

      Pixel scenes

      Living pixel scenes drawn entirely in code, from a Raja Ampat afternoon to Jakarta at 2 a.m.

      4 scenes, no image files

      Built with Canvas 2D and plain JavaScript on one shared animation loop.

      Open the pixel scenes
    • Open source

      Idea Wall

      An idea board that lives on the desktop wallpaper and syncs both ways with a Markdown notes vault.

      A new note shows up within seconds

      Built as a Node web app for the wallpaper, plus a native macOS app in Swift 6.

      Read the code on GitHub (opens in a new tab)

    Skill mapAgent HQ

    Contact
    Agent HQ

    Eighteen agents, one room each.

    As of September 2026, nine are running, three wait for a call, two are still being built and four are paused. Pick a room to see what it does.

    • 9 running
    • 3 on call
    • 2 in progress
    • 4 paused

    Click a room, or Tab into the building, move with the arrow keys and press Enter. The sky over the roof follows Jakarta time; it is there now.

    • Trading Command Center

      On call. Runs 12 paper-trading bots in one dashboard. Claude and Groq both have to agree before any trade; when they disagree, the bot holds.

    • Market Intelligence Agent

      In progress. Tracks big-fund filings, insider buying and factor scores, ranks the signals and writes a briefing. The build is done; it is not deployed yet.

    • XMON

      Running. Polls a list of X accounts, has Claude label each post by market direction, then forward-tests those calls against paper returns.

    • Content Automation Pipeline

      Running. Turns RSS news into short videos: an LLM rewrites the script, Whisper voices it, a renderer builds the clip and it posts to TikTok and Instagram.

    • Camera AI demo

      On call. Recognises me on camera, captions the live feed, transcribes what I say and hands spoken requests to an app-building agent.

    • Resume tailor

      In progress. Picks and orders resume bullets for a job description. A code gate blocks anything that is not in my real facts, and the whole engine is served over MCP.

    • Thesis RAG chatbot

      Running. Answers questions from a university's thesis guidelines, and screens every new document for poisoned or conflicting content before it enters the index.

    • Email Reactor

      Running. Sorts my inbox and sends a daily digest. Its client mode opens the right project, tries a fix on a branch and drafts a reply for me to check; that mode is paused for now.

    • ClaudeClaw

      Paused. A Telegram bot that answers each message by starting a fresh Claude Code session with the chat history. My first agent outside the terminal.

    • Threads MCP

      Running. Gives agents Threads as tools: read the feed, draft a post, publish it. Every write stays a dry run until I confirm it.

    • LinkedIn AutoPoster

      Paused. Drafted and published a LinkedIn post twice a day. A one-week build from the end of February that I have since paused.

    • Founder's life log

      Running. For a founder: transcribes voice notes, files each one under a part of life, and syncs wearable data every six hours.

    • B2B sales agents

      Paused. For a US client: agents that find leads, score them, draft outreach and brief the team before meetings. A demo I delivered in March.

    • Support agent demo

      On call. Reads a customer email, looks up the order with a tool call and drafts the reply, showing each step as it happens. Built to show clients.

    • Claude Code

      Running. My main agent and the engine behind every room: a written SOP, long-term memory, a daily log, custom skills and teams of subagents that split big jobs.

    • Mahoraga

      Running. Reviews every finished session, turns real mistakes into rules, and loads the locked rules into the next session so the same mistake does not happen twice.

    • PC Monitor

      Running. Watches my Mac's disk and memory, clears space before it runs out and takes remote commands over Telegram when I am away from the desk.

    • Amadeus

      Paused. A digital twin that watched what I worked on, woke on a two-hour pulse and reasoned in my voice over Telegram. Paused since June.

    Skill mapJourney

    Contact
    Journey

    From OpenClaw to agents of my own.

    I met my first AI agent at 01:50 one night in February 2026, and it ended in an error. These are the eight months since. Scroll the list and I walk the level: each flag is a month, each block holds something I learned, and the agents I built fall in behind me.

    The level follows the month you are reading.

    1. The night I met OpenClaw

      • At 01:50 I was trying OpenClaw, an open-source personal AI agent, and hit my first error with it.
      • Ran three Claude agents at once: one wrote files, one changed them, one checked the other two.
      • Picked Claude Code as my own agent and gave it OpenClaw's kind of memory: an SOP, long-term notes and a daily log.
      • Put an agent inside a Telegram bot, with a task queue and memory. It became ClaudeClaw.
      Unlocked
      • OpenClaw
      • Multi-agent
      • Memory
      • Telegram bots

      Agents born this month: ClaudeClaw, LinkedIn AutoPoster, Claude Code

    2. Agents leave my laptop

      • First client ship: a health platform, deployed and security-audited by a swarm of agents.
      • First MCP tools: Playwright over MCP, so an agent can click through the UI it just built.
      • Claude Code running headless on a VPS, signing people in with OAuth.
      • A sales-agent demo on the Claude Agent SDK, wired to Gmail, Calendar, Notion and n8n.
      Unlocked
      • Agent swarms
      • MCP
      • Headless on a VPS
      • Agent SDK

      Agents born this month: Trading Command Center, Content Automation Pipeline, B2B sales agents

    3. Agents that learn from mistakes

      • Designed Mahoraga: hooks catch my agent's mistakes and promote them into permanent rules.
      • One build day, one 95-file mobile app for a client, written by seven agents in parallel.
      • My first hook in daily use: it loads the right memory by keyword in about 230 ms.
      Unlocked
      • Self-adaptation
      • Parallel builds
      • Hooks

      Agents born this month: Founder's life log, Mahoraga

    4. Skills, and a twin

      • Four skills in one day, Seismic Sense, Auto-QA, Writing DNA and Code Janitor, published on GitHub.
      • Took the Hermes agent framework apart, six subsystems, and held each one up against my setup.
      • Started Amadeus, a digital twin that reasons from my own past decisions.
      • Mahoraga closed the loop: an LLM now reviews each session and locks the lessons.
      Unlocked
      • Skills
      • Digital twin
      • Closed loop

      Agents born this month: Market Intelligence Agent, Email Reactor, PC Monitor, Amadeus

    5. The twin goes live

      • Amadeus went live: three scheduled collectors feeding a model of how I think, reached through Telegram.
      • For my thesis, a RAG pipeline whose custom n8n node stops conflicting documents before they reach the index.
      • An always-on Claude companion squeezed onto a 1 GB VPS.
      Unlocked
      • Person model
      • RAG
      • Always-on companion

      Agents born this month: Thesis RAG chatbot

    6. Long runs and live demos

      • My first multi-hour autonomous loop: an agent that paced its own work over dozens of rounds.
      • A public support agent with real tool calls, live on a VPS.
      • ClipCraft, an open-source app built mostly by agents: 306 commits.
      Unlocked
      • Long autonomous runs
      • Tool use
      • Open source

      Agents born this month: Support agent demo

    7. Back to OpenClaw, from the other side

      • Scoped a job looking after a client's fleet of OpenClaw agents, the tool I started with in February.
      • Browser agents that work inside my own logged-in Chrome.
      • A game prototype built by isolated worker agents running in parallel.
      • Skill Watcher: measures what each skill costs in tokens with live A/B runs.
      Unlocked
      • Agent fleets
      • Browser agents
      • Cost tracking

    8. Agents of my own

      • XMON, a signal agent that labels posts from X by market direction and forward-tests the calls.
      • Rasi, an agent that starts from a blank page and thinks for itself. I only build its infrastructure.
      • A Threads connector, as a command-line tool and as an MCP server.
      • A resume-tailoring engine with its own MCP server and a code gate against made-up facts.
      Unlocked
      • Signal agents
      • An agent of my own
      • MCP servers

      Agents born this month: XMON, Camera AI demo, Resume tailor, Threads MCP

    That brings the line to 18 agents, and the level ends at their building. Meet them at Agent HQ

    Skill mapContact

    Contact

    Tell me what is broken, or what you want built.

    Not sure which one? Describe the problem and I will tell you if it is mine to solve. Reply within a day, Indonesia time.

    rafiimanggala3@gmail.comGitHubLinkedIn
    • Available
    • Remote
    • UTC+7
    // 05 · FAQ

    Before you reach out

    Yes, open to remote AI engineering roles and freelance builds: autonomous systems, agent orchestration, and full-stack products.

    Rafii Manggala

    AI agents that do the work, not just write the code.

    Stack

    • TypeScript
    • Python
    • C# / .NET
    © 2026 Rafii Manggala Japamel.Built with Next.js · the agents helped.