AI Stack

April 19, 2026 · View on GitHub

A ~60 minute voice note walking through my current AI tooling, workflows, what's working, what isn't, and where the gaps are. Transcribed with Gemini (via the gemini-transcription MCP, raw mode), then broken down into structured sections.

Not a polished essay or a recommendation — a footprint: what I was using, what I was stuck on, and what felt like an opportunity on a specific Sunday during a fragile ceasefire.

📄 Full document

ai-stack-2026-04-19.pdf — sections + cleaned transcript as appendix.

Stack at a glance

Layer	Tool(s)
Voice capture (mobile)	ASR / Voice Notes → webhook
STT (desktop, daily)	Multimodal Typer + Parakeet
Transcription (long-form)	Gemini multimodal (single-pass format)
LLM (agent work)	Claude Code (Opus 4.7)
LLM (script / creative)	Sonnet
LLM (routing / experiments)	OpenRouter
Image gen	Nano Banana 2 via Fal
TTS	Chatterbox on Modal (voice-cloned)
Memory	Mem0
Vector / RAG	Pinecone
MCP hosting	Home server + MCP Jungle aggregator
Remote access	Cloudflare Tunnel / Tailscale
Deploy	Vercel
Domains / edge	Cloudflare
Backup	NAS + Wasabi/S3
Doc output	Typst
GPU (on demand)	Modal

Organising pattern — repo-as-workspace. The utility of a repository is the folder structure, not version control. Standard layout: prompts/, outputs/, context/. Skills and MCPs layer on top.

Repo layout

.
├── audio/                      # source recording (VAD-trimmed)
│   └── 2026_04_19_13_01_40.opus
├── sections/                   # structured notes extracted from the voice note
│   ├── stack.md                # layered view of the working stack
│   ├── tool-list.md            # every tool mentioned, grouped by function
│   ├── thoughts.md             # philosophy, takeaways, unresolved questions
│   ├── blockers.md             # concrete pain points and friction
│   └── opportunities.md        # gap-shaped ideas worth exploring
├── transcripts/
│   ├── transcript-verbatim.md  # raw Gemini transcription, unedited
│   └── transcript-cleaned.md   # readable paraphrased pass
├── document.typ                # Typst source for the compiled PDF
├── styles.typ                  # shared Typst styling
├── build.sh                    # regenerate the PDF
└── ai-stack-2026-04-19.pdf     # compiled output

Source audio

audio/2026_04_19_13_01_40.opus — ~60 min, recorded on Android, silence trimmed via ffmpeg VAD (-30 dB RMS, 0.8 s min-silence).

Rebuilding the PDF

./build.sh

Requires pandoc and typst.