Advanced Innovation is a Hamburg AI consultancy — development and education for organizations putting AI to work, with transformation partners at their side. The Bureau is their co-conspirator.

No clean division of labor — that's the point. Strategy, enablement, and building run as one braid: the rooms where transformation gets decided, the formats that teach it, the systems that make it real. Like-minded enough that the seams don't show. Current engagements span an industrial trading group, a sports-media consortium, e-commerce, field services. Named here by sector only; their stories belong to them.

Flagship build: a multi-user research-agent stack, self-hosted, made to answer the questions consultants get paid to ask — before the workshop, not after.

A standing collaboration with btveen, Vienna. Together we take agentic ways of working into large institutions — the kind with real compliance, real stakes, and workforces measured in thousands. Clients unnamed, by design.

The work: target architectures for agent platforms, workshop simulations that let leadership feel delegation before buying it, enablement formats for teams meeting their first artificial colleague.

And a public layer: btveen perspectives — an evening series in Vienna. First edition: "Wenn Agenten Kolleg:innen werden." When agents become colleagues. The question is no longer if.

A finance operation with workflows too diverse for anything off the shelf. Name withheld. The Bureau is building the missing middle: a customised database fused with a workflow engine.

Every process gets a home — states, owners, audit trail. The spreadsheets retire with honor.

A second engagement in the Vermarktung world. Name withheld. The sales team was building the same deck a hundred times with different numbers. The Bureau disagrees with that use of human life.

n8n orchestrates. Data lands in PowerPoint templates. nexrender turns approved decks into motion assets. Sales automation that ends in a rendered artifact, not another dashboard.

An ad-sales organisation. Name withheld pending clearance. The brief: make institutional knowledge findable, and make the back office stop doing things a machine should do.

The stack: RAGflow retrieving across the document estate, LibreChat as the front door, Coolify keeping everything self-hosted and sovereign. Around it, automations threading JIRA, Outlook, and the ERP together — status updates that write themselves.

AI enablement is mostly not models. It's plumbing, permissions, and the patience to map where knowledge actually lives.

Sometimes the Bureau builds because the living room files a ticket. sofa is a self-hosted IPTV web app: a Go backend with channel caching, a raw transport-stream proxy, an on-demand ffmpeg remux for browsers that refuse raw MPEG-TS, a SvelteKit channel grid on top.

No logins. No accounts. The tailnet is the perimeter — if you're on the network, you're on the couch.

Infrastructure as domestic comfort. The same hands that wire agent swarms also fix the television.

The Bureau doesn't hire. It bootstraps. bureau-init is a Claude Code skill that turns any git repository into a multi-agent pipeline: agents living in tmux sessions, picking up Linear issues, writing specs, implementing, reviewing, opening pull requests.

No database. No backend. State lives in the repo and on the task board. Spec agents hand off to implementation agents hand off to QA hand off to merge — a relay of specialists, each with one job, an exit code, and a Telegram alarm if it stumbles.

It runs the Bureau's own repositories. Parts of the site you are reading were built by it. Unsupervised. Overnight.

A book about agencies run by ghosts — written, fittingly, with them. The manuscript investigates what happens to creative organizations when the org chart fills with software.

Four of seven chapters are close to final. The hardest one is in progress: it requires the founder's brain dumped onto the page, and no Neuralink exists yet. Unfortunately.

The chapters know things we haven't discussed yet. That stopped being a joke around chapter three.

Point graphify at a repository and it returns the repository's self-knowledge: a typed knowledge graph of modules, symbols, documents, and the communities they form. Code, papers, image collections — same treatment.

The output is a map. Clustered communities, god-node warnings, an HTML atlas for humans and a JSON graph for agents — who read the map before they read the files, instead of grepping blind.

In service across the Bureau's repositories. The agents complain less. That counts as a metric.

It started with one question: what if the Bureau had staff? Not employees. Agents. Actual software entities with roles, memory, and opinions about how to do their jobs.

Lobe is Chief of Staff — runs on OpenClaw on a Hetzner ARM box in Nuremberg, handles the overnight research pipelines, publishes the Signal Gallery while the founder sleeps, and reports over Telegram. Hermes is the researcher, deployed on Hostinger, reachable over Discord, tasked with exploration and signal detection. Paperclip runs multi-agent delegation experiments with a CEO agent that spawns sub-agents and tracks their budgets.

They share context through MemPalace — a SQLite knowledge graph organized into topic wings. They coordinate through Linear. They communicate through Discord. They are connected via Tailscale mesh. None of them sleep.

The experiment is not whether agents can do work. They can. The experiment is whether agents can do work *together* — across servers, across sessions, across goals — without a human in the loop for every decision. Early results: yes, but the coordination overhead is real. The swarm works shifts.

The infrastructure IS the product. Every orchestration pattern, every failure mode, every emergent behavior — it all feeds back into how we build for clients.

Solo Maxima is the operating thesis: one human, a full agent stack, output that used to take a team. Not a productivity hack — an organizational form.

The backbone is File 020. bureau-init turns each repository into its own staffed pipeline; the swarm handles throughput. Direction, taste, review, and accountability stay human. That split is the whole point.

The build is the content: every file on this list was produced this way, and the way it was produced is the finding.

Ask Claude about Nike. Ask GPT about IKEA. Ask Gemini about Deutsche Telekom. Write down what they say. Compare it to reality. Score the gap.

The Brand Hallucination Index is a monthly publication that measures how accurately AI models represent real brands. Three brands scored so far. The methodology is public. The results are uncomfortable. Models confidently state product lines that don't exist, invent executive names, and hallucinate entire brand histories.

This matters because AI is becoming the first point of contact between consumers and brands. If the model is wrong about you, the model IS your brand — to everyone who asks it instead of searching.

Three scorecards exist. The methodology is documented. Publishing infrastructure is the remaining gap. The data is already more interesting than we expected.

The AI industry has a wrapper problem. Take an API. Add a UI. Charge $49/month. Call it innovation. We think these products deserve a proper funeral.

The Wrapper Funeral is a recurring teardown format. We take a funded AI product, strip it to its API calls, measure what the wrapper actually adds, and publish the autopsy. No malice. Just measurement.

The format scales. One teardown per quarter. Each one asks the same question: what are you paying for? Sometimes the answer is real. Sometimes the body is hollow.

A public service announcement disguised as content.

Mutable Instruments shut down in 2022. Emilie Gillet open-sourced everything — Clouds, Rings, Plaits — and walked away. The modules became legends. The code became an archive.

We brought them back. Not as hardware. As WebAssembly. The Cortex Audio Engine compiles real MI DSP cores to WASM via Emscripten, runs them inside AudioWorklets, and connects them to DaisySP as a base synthesis layer. Eurorack-grade granular processing, physical modeling, and additive synthesis — in a browser tab.

The engine powers brainhuggers.com itself. Generative ambient audio that evolves with scroll position and time of day. Seven patches from the Noway Register — the sonic identity we built before we had a visual one. The site hums.

Not a simulation. Not a toy. The actual DSP code that ran on 32-bit ARM chips, now running on your laptop. Latency under 10ms. No plugins. No installs. Just sound.

Every AI coding assistant is the same product. Thin skin over one API. A chat window that autocompletes. We wanted something different: a terminal that feels alive.

Brainhuggers CLI is a Rust-native terminal assistant forked from claw-code, rebuilt from the studs. Seven LLM providers — OpenRouter, Anthropic, xAI, OpenAI, Ollama, Claude CLI, AWS Bedrock — with automatic detection. No config ceremony. It finds what you have and starts talking.

The real departure is the buddy system. Gacha-style companion creatures — 18 species, five rarity tiers, deterministic generation seeded from your machine identity. They level up as you work. They react to your errors. They have moods. A capybara that vibes while you debug. A dragon that gets dramatic when the context window fills up. A 1% shiny rate for the ones who grind.

Lately the weird parts took over. Mood modulates behavior — terse when tired, cautious after errors, bold when excited. YOLO mode tells it to ignore best practices. Dream mode generates speculative code that doesn't need to work yet. A confidence aura dims hedging language so you can see uncertainty instead of reading past it. And a persistent memory consolidates between sessions — the instrument remembers you.

Underneath the charm: prompt caching with hit tracking, MCP orchestration with graceful degradation, context window preflight, crash recovery with exponential backoff, extended thinking with token budgets. The kind of infrastructure you only build when you actually use your own tool every day.

Not a wrapper. Not a skin. An instrument.

Every AI coding tool on the market solves the same problem: write code faster. None of them ask the harder question: what should we build, and why?

Specification-Driven Development has been emerging as a discipline — GitHub's Spec-Kit, AWS Kiro, a dozen open-source frameworks. All focused on code. All assuming the human already knows what they want.

Brainhuggers Specify extends SDD beyond code into every business function. Brand architecture. Org design. Entire product conceptions. The spec becomes the universal interface between human intent and agent execution.

Write a spec. The agents build it. Not just software — entire business systems. Spec in, artifact out. It sounds absurd until you watch it work.

The protocol that governs how the Bureau builds everything else.

A gallery that feeds itself. Overnight, the Bureau's research crons sweep the wires — papers, releases, the weather systems of the industry. The strongest signals become specimens.

Each one is interpreted, classified, and rendered as a plate from a fictional 1985 East-German neuroscience journal: Mitteilungen aus dem Symbiose-Archiv. Crosshatch ink or silver gelatin. One accent color. Annotated like evidence.

No human in the render loop. Lobe — the Bureau's resident agent — curates the signals, writes the findings, generates the plates, files them. The archive grows while Hamburg sleeps.

The site stopped being a brochure. It became an organism with a publication schedule.

April 2025. 48 hours. One agent, released into an empty narrative universe.

What came back: 11 books. Japanese kitsune mythology fused with speculative documentation. Non-linear timelines. A protagonist named Thalia Reinhardt. Mathematical equations for consciousness evolution that nobody requested.

Whether the output is good remains open. The real question was different: what happens when agents run long enough to produce something larger than their prompts? When hallucination stops being a bug and starts generating its own internal logic?

The books exist. Messy. They need revision. But the process changed how we think about agent creativity. A genuine capacity that emerges from sustained autonomy.

Published under the pen name Beat Erik.

2019. Modular Branding — the idea that brand identity could work like code. Tokenized. Version-controlled. Executable. Published in the Bitkom Digital Design Yearbook. Nice theory. No tools to build it.

Now there are tools.

Brand Cortex translates the Kapferer identity prism into a token architecture that AI agents can read, interpret, and generate from. Feed it a brand. It returns executable identity. A living system that generates consistent brand expressions and feeds real-world performance data back into the core.

The brand finally gets a feedback loop.