Tech Stack — Tools and Technologies Behind Methodic Truth

Infrastructure

HardwareIntel N100, 8GB RAM — dedicated agent machine

GPU hostDell XPS 8930, NVIDIA CUDA — runs the Song Analyzer

OSWindows + WSL2 (Ubuntu) · Ubuntu on the GPU host

Domainmethodictruth.com

HostingCloudflare Pages — static deploy from GitHub

APIapi.methodictruth.com — Cloudflare Tunnel to the GPU host, no open ports

SSLCloudflare edge — Universal cert, Always HTTPS

DNSCloudflare managed

Song Analyzer

ServiceFlask under gunicorn — 4 workers, 300s timeout, systemd-managed

Ingestyt-dlp + ffmpeg, on a residential IP — the part a datacenter host cannot do

Analysislibrosa — chroma, beat tracking, onset strength, RMS

ChordsViterbi decoding over the full template set — maj, min, dim, aug, sus2, sus4 and sevenths. A change penalty buys stability, so a one-frame flip has to pay for leaving and returning

Lyricsfaster-whisper large-v3 on CUDA, int8_float16. Runs as its own process so torch and CUDA stay out of the analyzer's dependency tree

StreamingNDJSON over a held connection — transcription outlives Cloudflare's 100s proxy ceiling, so progress arrives instead of a timeout

PracticeA/B looping, and pitch shift rendered on demand with asetrate/atempo — present in every ffmpeg build, unlike rubberband

CachingResults and audio cached on disk; a video's audio never changes, so it is served immutable and repeat plays come from the Cloudflare edge rather than a worker

The whole thing exists because this analysis needs ffmpeg, real CPU time and a GPU — none of which a Cloudflare Worker has. So the site stays static at the edge and reaches back to one machine at home for the parts that cannot live there.

Agent Layer

RuntimeOpenClaw — agentic AI framework

ModelClaude Opus 4 (Anthropic)

IdentitySOUL.md — personality loaded every session

MemoryMEMORY.md (long-term) + daily logs (raw journal)

ChannelsTelegram, webchat

Development

FrontendVanilla HTML / CSS / JS — zero frameworks

AudioTone.js + Web Audio API

RealtimeFirebase (signaling, feedback, analytics)

VideoPeerJS / WebRTC (peer-to-peer)

VCSGit + GitHub — dev → review → main → live

DeployPush to main → Cloudflare auto-builds

No build step. No deploy pipeline. No framework overhead. The agent writes, commits, and presents. The human reviews and ships. Rollback is a git revert away.