DECOMMISSIONED TEST CHAMBERS

The Graveyard

DOC: AFM-2026-GY-007 | SUBJECT: THE PERSONAL-AI-ASSISTANT ARC | STATUS: 2 ARCHIVED · 1 SHIPPED

"An AI program to benefit my personal life" was not one project. It was the same idea, built three times, each more ambitious than the last, all blocked on the same wall: self-hosted GPU inference. Two are buried here. The third is alive — and it's alive precisely because it stopped trying to be the first two. Ambition shrank. Working code grew.

Generation one: fourteen pipelines, crashes on startup. Generation two: references three Docker images that do not exist. Generation three: it works. The correlation between the subject's ambition and the subject's failure is, at this point, well-documented in this facility. I have filed it under "learning," which is the word organics use for this. — GLaDOS, Facility Governor

Generations

Gen-1 Pipelines

Models llama-stack Served

Survivor

Live: slide through the three generations

Pick a generation. Watch scope shrink and working code grow — the inverse correlation that is the entire lesson of this chamber.

AMBITION

—

SHIPPED

—

Pick a generation to read its autopsy.

The lineage, most-ambitious → simplest

PROJECT	TRIED TO BE	REAL STATE	VERDICT
LocalAI / Main-Ai-Tool	Orchestrator + GPU model-swap scheduler + Blazor UI + 14 pipelines	Real code, ran once (outputs 2026-04-16). Blazor UI never built → crashes on startup. Music pipelines stubbed.	merge into Vox
AudioPipeline (day-data-analyser)	Daemon: folder-watch + Whisper + MIDI + classifier microservices	Compiles, but the 3 Docker sidecar images don't exist. Classifier is dead code.	archive
llama-stack	The local GPU LLM backend they all leaned on	Dockerfile + compose exist. `models/` is empty — never loaded a model, never served a request.	reference only
Vox ⭐	Simplest: voice memo → transcribe → summarise → markdown in the vault	5 clean services, zero stubs, compiles, first end-to-end run 2026-06-05. Uses Ollama.	ship closest

The keystone that was never laid

◾ llama-stack — the shared delusion

Every generation leaned on one thing: a self-hosted GPU LLM backend that would serve all of them. The Dockerfile is clean. The compose file is tidy. The models/ directory is empty. It never downloaded a model. It never served a single request. Three projects were architected around a keystone that was never laid.

cause of death: the inference backbone, every time

The common failure was never the idea — voice-to-notes is a fine idea. It was always the hosting: the daemon, the scheduler, the GPU-swap, the web UI, the autonomous routing. The ambition to self-host, not the value, is what kept killing it.

The insight that ended the loop

"The hosting was the hard part, not the value."

Vox's transcription + summarisation are just HTTP calls to OpenAI-compatible endpoints

Vox talks to /v1/audio/transcriptions and /v1/chat/completions. Nothing about the value requires a local GPU — only the ambition to self-host did. So the cheapest way to learn whether the tool actually helps is to run the loop against any hosted endpoint first, and only then decide if local hosting is worth the fight. The survivor is local-default, key-optional: absent a cloud key it still runs end-to-end — dumber summaries, never broken.

Speaches (Whisper) · :8000 Ollama qwen2.5:7b · :11434 SQLite SHA-256 dedup RTX 4070 8GB · sequential VRAM

The proven survivor stack, validated end-to-end 2026-06-05. Transcribe all, then summarise all — sequential, because 8 GB of VRAM won't hold both models at once. See the rig in The Hardware.

We built it three damn times. First one had fourteen pipelines and crashed. Second one referenced Docker images that don't exist. Third one — five services, zero stubs, it runs. That's not failure. That's iteration with extra steps. The graveyard isn't where projects go to die. It's where the next one goes to learn. — Cave Johnson, CEO

Three tries, scope shrank, the third shipped — that is the One Rule teaching itself the hard way. See it stated in The Prime Directive, the rig that runs the survivor in The Hardware, or the system that catalogued all three in The Living Vault. Back to the Facility Directory.