I build systems whereAI agents do the work,not just write the code.
AI Engineer · Autonomous Systems. I deploy Claude Code as the engine: always-on agents that observe, decide, and act in production: trading bots, digital twins, and autonomous infra.
- 15
- Always-on AI agents
- launchd, 24/7
- 59
- Self-learned rules
- auto-locked, permanent
- 94
- Daily build logs
- shipped + logged
- 39
- Git repositories
- 81 project folders
- 6
- Autonomous systems
- running in production
- 2
- Client countries
- AU · US
Watch an agent ship a fix.
A real loop my agents run unattended: read the issue, fix it on a branch, deploy, verify with Playwright, attach proof. Press run.
I run AI agents in production, not in a chat window
I operate a fleet of AI agents as a working system: 15 always-on processes, 6 autonomous pipelines, and a self-learning rule set that ships client work while I sleep. I build the orchestration, the safety rails, and the tooling myself, then drive all of it from my phone.
Before coffee, an hourly Email Reactor has already caught a client bug report, spun up a headless Claude, warmed up the right project, and fixed it on a branch waiting for review.
Mid-morning, a single task fans out into 5 to 12 parallel subagents: one writes code, others research and debug, a lead agent merges the results.
An adversarial-verify workflow runs a judge panel over the work, looping until every check passes instead of trusting the first answer.
A TE loop closes a feature end to end: Fix, Deploy, Test through Playwright, then send proof screenshots before anything is marked done.
At session end, Mahoraga reviews the transcript and locks a confirmed lesson into a permanent rule, so the same mistake never recurs.
All of it is steerable from an iPhone over Termius and Tailscale SSH, from anywhere, at any hour.
Questions
How does Rafii Manggala use AI?
Rafii runs Claude Code as an always-on personal agent operating system on a Mac Mini M4, not as a chatbot. Fifteen launchd agents run around the clock, six autonomous systems run in production, and he steers the whole fleet from his iPhone over SSH. AI is his production environment, not a tool he occasionally prompts.
What makes Rafii's AI workflow different?
He builds the orchestration himself: self-learning hooks that lock in lessons permanently, an email reactor that fixes client bugs autonomously, parallel agent teams with adversarial review, and his own desktop and web automation tooling. Most people use AI to get answers; Rafii engineers AI systems that ship work, with safety rails and proof loops built in.
Can Rafii ship production work with AI agents?
Yes. He has delivered client work for health and education platforms (BioBrain on .NET, HodieLabs) across Australia and the US, with every fix Playwright-proven before he reports it done. His TE loop deploys to an isolated test port, runs the change through a real browser, and attaches proof screenshots, so done means verified, not claimed.
Not a chatbot. An operating system for agents.
The part most engineers do not have: a personal agent stack that runs itself, learns from its own mistakes, and never touches production without a gate.
Mahoraga
Self-adaptation system that captures the agent's own failures and promotes real lessons into permanent rules.
A genuinely closed-loop self-improving agent: an LLM reviews each session transcript and auto-locks confirmed lessons. 30+ locked patterns.
Email Reactor
Reads new client email, warms up the right repo, fixes the issue on a branch, and drafts a reply.
A self-driving freelance assistant: spawns headless Claude per inbound email, stops just short of anything irreversible.
Smart Context Injector
Auto-loads relevant project memory by keyword match. Pure bash, ~0 token cost, 230ms.
Two-layer retrieval: instant free keyword injection + on-demand semantic search. Persistent memory, no token overhead.
Parallel agent teams
Lead agent delegates to a fleet of teammates: inventories, audits, and full-stack builds run on 3-5 concurrent agents.
Multi-agent council review
A 4-agent adversarial panel that must all pass before work is 'done'. Adversarial self-verification built into the workflow.
Self-adaptation hooks
Closed-loop learning: LLM transcript review → tentative → confirmed → locked rules, reloaded every session.
The tools I built to build faster.
15+ skills, hooks, and CLIs wrapped around Claude Code. The interesting work isn't prompting, it's the harness around it.
$ hooks/mahoragaHook systemMahoraga
Self-adaptation system that captures the agent's own failures and promotes real lessons into permanent rules.
A genuinely closed-loop self-improving agent: an LLM reviews each session transcript and auto-locks confirmed lessons. 30+ locked patterns.
$ /graphifyCLIGraphify
Turns any input (code, docs, papers) into a clustered knowledge graph, self-rebuilding on every commit.
Deployed across 6 live projects; one codebase graphed at 27k nodes / 36k edges, queryable in natural language.
$ /seismicSkillSeismic Sense
UI/UX structural analysis without screenshots, extracts design DNA and scores live pages deterministically.
Replaces token-heavy visual review with structural analysis: ~88% token savings, zero LLM tokens on scoring.
$ /dnaSkillWriting DNA
Learns a person's writing voice from real messages and generates new text that sounds like them, per recipient.
Self-correcting via a feedback log; powers autonomous email drafting that passes as the real author.
$ /cua-driverInfraCUA Driver
Drives native macOS apps by accessibility tree without moving the cursor or stealing keyboard focus.
Solves the 'macOS has one cursor since 1984' problem via SLEventPostToPid: the only non-VM path to true cursor isolation.
$ wa auto <url>CLIwa
4-tier stealth web-automation dispatcher for anti-bot targets (Cloudflare / Turnstile / DataDome).
Built on the 2025-26 anti-bot meta: curl_cffi + nodriver (raw CDP), with compile-to-script for cheap token-free reruns.
$ hooks/memory-injectHook systemSmart Context Injector
Auto-loads relevant project memory by keyword match. Pure bash, ~0 token cost, 230ms.
Two-layer retrieval: instant free keyword injection + on-demand semantic search. Persistent memory, no token overhead.
$ launchd · hourlyInfraEmail Reactor
Reads new client email, warms up the right repo, fixes the issue on a branch, and drafts a reply.
A self-driving freelance assistant: spawns headless Claude per inbound email, stops just short of anything irreversible.
Orchestration, not autocomplete
Selected UI/UX work
Two case studies that show the process, not just the result: the problem, the flow, the system, and the screens.
Systems that run themselves.
Four builds where the AI isn't a sidekick, it's the engine making decisions in production. Click a card for the deeper version. Client work is anonymized.
Trading Command Center
Unified paper-trading platform running 12 AI + algorithmic bots under one dashboard, gated by multi-model consensus.
- ▸Multi-model consensus gate: Claude and Groq must agree before any trade. Disagreement forces HOLD. Research-backed anti-hallucination design.
Amadeus · Digital Twin
An always-on agent that observes my activity, learns my cognitive patterns, and reasons in my own voice.
- ▸Decision-exemplar bank: mines 320 real decisions from my git commits + self-correction logs, retrieved via BM25 so the twin reasons 'as I would' instead of guessing.
TestEngine
Self-hosted MCP server for isolated, parallel browser testing, one Docker container per session.
- ▸27 MCP tools across session / browser / auth / pool / debug; container pool with pre-warm, recycle, and a health monitor enforcing memory/uptime limits.
Health Optimisation Platform
Health web app ingesting biomarkers, DNA, DEXA scans and wearables into clinical scores + AI recommendations. (AU client)
- ▸DEXA body-composition parsing across scanner vendors, with auto-crop of report images via PDF operator-list inspection and reference-based percentile scoring (not hardcoded).
I let Claude drive the simulator.
React Native, .NET MAUI, SwiftUI. The agent boots the simulator, builds the screen, screenshots it, fixes what's off, and verifies, while I review the architecture.
More things I've shipped.
A sample of the other 40-odd repos. Hover a row for the stack.
Autonomous & AI systems
- Market Intelligence Agent13F whale + Form-4 insider + factor signals into briefingsFastAPI · Anthropic SDK · SEC EDGAR
- ClaudeClawMulti-agent Telegram bot (Haiku UI + Opus workers)Node · Haiku/Opus · Groq
- Auto-Approve / PC MonitorHeadless Mac controlled fully from iPhone over Telegramlaunchd · Telegram · AppleScript
- Personal Assistant BotTelegram twin: voice transcribe + life-category classify + WHOOPNode · Claude CLI · Groq
- AI SDR Platform3 sales agents on a VPS for finance/law firms (US client)Agent SDK · Express · PG · MinIO
Full-stack products
- K-12 Education SaaSCurriculum learning platform + AI tutoring, 995 schools (AU client).NET 9 · Angular · PostgreSQL
- Employee Wellbeing AppPsychosocial-risk mapping + ISO 45003 reports (AU client)React Native · Express · MongoDB
- AI Surf Forecasting200+ spots, 16-day forecast, 'Surf DNA' matching (AU client)React · Express · MongoDB
- open-wearablesWearables data full-stack + MCP, full CI/CDReact/Vite · FastAPI · MCP
- Shopify Fashion StoreTheme fixes + full Klaviyo email program, 10 live flows (AU client)Liquid · Klaviyo · Shopify CLI
Tools & infra
- Idea WallNative macOS menu-bar idea board, zero third-party depsSwift 6 · WKWebView · Claude CLI
- Code JanitorDead-code + semantic-clone scannerKnip · jscpd · LLM layer
- Auto-QA + WrapupPost-codegen quality gate + cross-session bannersHooks · Seismic · figlet
- Bloomberg / MMT terminalsCommand-driven financial terminalsFastAPI · Alpine · uPlot