15 September 2026 · Kosmos Berlin · Germany

Serverless AI Factory:
An Autonomous Dev Pipeline on AWS

Telegram → plan → code → PR · Hostinger + OpenClaw · HalloCasa · ai-secure · qr-plakat on AWS

aws-community-day.de/agenda

Speaker

Martin Mueller

🇵🇹 🇩🇪 Moved back to Germany · February 2026

Freelancer · DevOps · AI agents

OpenClaw on Hostinger VPS · Telegram as control plane
Manages AWS: HalloCasa · ai-secure.dev · qr-plakat.de
Pipeline with human gates: issue → PR → you merge

linkedin.com/in/martinmueller88 · martinmueller.dev

Mueller family baking Christmas cookies, 2025

Family · Christmas 2025

Context

Clients + AWS deployments I manage

Until July 2026

Air Canada

Enterprise AI chatbot · customer service · conversational AI in production

Current

HDI Versicherung

Cyquins cyber insurance · automated cloud risk scoring on AWS

Own project · AWS

HalloCasa

CDK · CodePipeline · prod SSO — real estate platform

Own project · AWS

ai-secure.dev

AgentCore · MCP scans · compliance reports (ISO, SOC 2, NIST)

Own project · AWS

qr-plakat.de

OpenNext · Lambda · CloudFront — hosted landing pages with metadata preview

Freelance

Open for work

AWS · DevOps · AI agents — book 15 min if you need help

Calendly

calendly.com/martinmueller_dev

Right now

What excites me: ChatGPT Ads

Sponsored martinmueller.dev ad card — AWS ohne Overhead — The ad · martinmueller.dev

ChatGPT conversation showing the sponsored placement in context — In context · live ChatGPT placement

Distribution at intent — sponsored placements below answers, matched to conversation context (not a search results page)
Platform maturing fast — Ads Manager in Europe · CPC + conversion measurement · still early formats
Builder angle — product feeds, landing pages, attribution — same craft as growth + infra work
Watching it like early Google/Facebook ads: messy now, obvious in hindsight

Personal interest, not a pitch — happy to swap notes after the talk if you're experimenting too.

Agenda

~30 minutes

Speaker intro · clients · what excites me (ChatGPT Ads)
What I actually run — VPS + Telegram + AWS
Human in the loop — not full autonomy
Dev pipeline flow — issue → PR
Two-tier architecture — control plane vs AWS targets
Managing AWS from Telegram — SSO, pipelines, ops
Economics — VPS + AWS + LLM spend
Live demo — Telegram → agent → PR
Practical learnings — CI, cost, models, dual track

Honest architecture

AI Factory — cool concept, skeptical for code

AI factory diagram — issues in, PRs out, human review at the merge gate

AI Factory — issues in, PRs out, 24/7. Great buzzword · I'm not trusting full autonomy on production code yet
Telegram — talk to the agent from your phone, anywhere
Hostinger VPS — OpenClaw in Docker, ~€10/mo, always on
Watches GitHub issues, plans, codes, opens PRs · AWS targets (HalloCasa, ai-secure, qr-plakat)
Human in the loop — review diff, SSO approve, kill runaway jobs. Non-negotiable when code matters — at least for now

Factory runs overnight · you own the merge. Telegram on the go — or pair deeply in Cursor IDE.

Agent runtime

What is OpenClaw?

Open-source gateway — self-hosted on your VPS (Docker), always on
Messengers in — Telegram, Slack · tools out — shell, GitHub, AWS SSO, email
Persistent memory, cron heartbeats, skills — an agent harness, not just a chat UI
Routes to Cursor CLI / LLMs · plans, codes, opens PRs — with human gates

Alternatives: Hermes (Nous Research · self-hosted personal assistant) · LangDock (hosted EU platform · DSGVO-oriented · less DIY)

Design principle

Human in the loop — always

Agents execute — plan, code, test, open PR
Humans decide — review diff, approve merge, kill runaway jobs
Full autonomy sounds cool · production needs guardrails
HalloCasa · ai-secure · qr-plakat: agent proposes · you SSO-approve + merge

Autonomous pipeline ≠ autonomous production. The factory runs 24/7 — you still own the merge button.

Hook

What if your dev pipeline never slept?

Telegram ping at 2am — or GitHub issue lands overnight
Agent on VPS plans, codes, opens PR — you review over coffee
Needs AWS? HalloCasa pipeline · ai-secure scan · qr-plakat deploy — you SSO-approve on phone

Overview

Pipeline flow

GitHub issue → Plan agent → Code agent → Test / review → Pull request → Human review & merge

TODO: real git history screenshot

Architecture

Two tiers — VPS control plane, AWS production

Telegram → OpenClaw on VPS → manages HalloCasa · ai-secure · qr-plakat on AWS

Key components

VPS: OpenClaw gateway · Docker · cron heartbeats · MEMORY.md
Interface: Telegram topics · Slack mirror · GitHub via gh CLI
Coding: Cursor CLI — plan on Opus, implement on Composer
AWS: SSO device-code · CDK deploy · CodePipeline trigger
Memory: markdown files on disk — MEMORY.md, skills, daily notes

Dev pipeline agents

Plan agent — reasoning model breaks issue into steps
Code agent — implements, runs tests, fixes failures
Review agent — automated sanity check before PR
Human gate — diff review, merge approval, token/spend caps
Persistent VPS shell · long Cursor CLI sessions · Telegram always on

Memory & state — why VPS helps

MEMORY.md + daily notes — curated long-term context
heartbeat-state.json — cron checks, issue polling
Skills + TOOLS.md — how-to for AWS SSO, gh, MCP

AWS learnings

Managing prod from Telegram

HalloCasa — CDK, CodePipeline trigger, prod DB ops via SSO
ai-secure.dev — AgentCore, MCP scans, compliance reports
qr-plakat.de — OpenNext on AWS · hosted landing pages with metadata preview
Agent reads TOOLS.md → knows device-code SSO flow → sends you URL + code
Real DevOps value: one interface (Telegram) across multiple AWS accounts

Economics

Real numbers

~€10

Hostinger VPS / month

varies

HalloCasa · ai-secure.dev · qr-plakat.de

~95%

LLM spend (subscriptions > API keys)

Subscriptions + fast CI keep the agent affordable at scale

Live demo

Issue → PR in minutes · merge when you're ready

1. Message in Telegram (or GitHub issue lands)
2. OpenClaw on VPS plans + Cursor CLI implements
3. PR opened — needs AWS? SSO device-code to your phone
4. CodePipeline / CDK deploy to prod
5. You review the diff and merge — human in the loop

Fallback: pre-recorded demo clip if live fails

Learnings

What actually moved the needle

CI/CD + GitHub checks — offload deterministic work from the agent (lint, typecheck, tests). Keep build times short — minutes cost money.
LLM subscriptions — Cursor Pro etc. instead of raw API keys where possible. ~95% of spend is tokens anyway.
Cheap models are enough — Composer, Grok for implementation; reserve Opus for planning. Good enough, much cheaper.
Dual track — ship from Telegram (async, mobile) and from Cursor IDE (deep focus). Same repos, same PR workflow.
Self-update — configure OpenClaw to upgrade itself + skills (gateway, clawhub, npx skills update) so the agent stays current without SSH babysitting.
Share the agent — give co-founders / colleagues access (Telegram group, Slack channel). Non-technical teammates get the benefit too — and it saves you being the bottleneck.

Questions?

Martin Mueller

martinmueller.dev

github.com/mmuller88 · linkedin.com/in/martinmueller88

aws-community-day.de/agenda

Serverless AI Factory:An Autonomous Dev Pipeline on AWS

Speaker

Martin Mueller

Context

Clients + AWS deployments I manage

Right now

What excites me: ChatGPT Ads

Agenda

~30 minutes

Honest architecture

AI Factory — cool concept, skeptical for code

Agent runtime

What is OpenClaw?

Design principle

Human in the loop — always

Hook

What if your dev pipeline never slept?

Overview

Pipeline flow

Architecture

Two tiers — VPS control plane, AWS production

Key components

Dev pipeline agents

Memory & state — why VPS helps

AWS learnings

Managing prod from Telegram

Economics

Real numbers

Live demo

Issue → PR in minutes · merge when you're ready

Learnings

What actually moved the needle

Questions?

Serverless AI Factory:
An Autonomous Dev Pipeline on AWS