Turn transcripts into structured docs by mixing cheap and capable LLMs

What it does

Feed it a transcript, meeting notes, or unstructured text. Lumentis asks a few questions about themes and audience, generates an outline for you to curate, then writes up detailed pages with varied formatting. The output is a clean, deployable project folder—no leftover scaffolding beyond a .lumentis directory that transparently logs every AI prompt and response.

The interesting bit

The tool treats LLMs like a budget-conscious film director: cast an expensive model for the intellectually demanding outline work, then switch to a cheaper one for the bulk writing grind. It estimates the cost before each run, so you know whether your Feynman Lectures recreation will cost 38 cents or four dollars.

Key highlights

• Model mixing: assign different LLMs to different stages (e.g., GPT-4o for structure, Gemini Flash for bulk writing).
• Upfront cost estimates before you commit to a generation run.
• Resumable sessions: interrupt and restart without losing your place or answers.
• Transparent operation: every AI message and state change is stored in the local .lumentis folder.
• Vercel-ready output: the generated site is essentially a clean static project, ready to push and deploy.

Caveats

• Returning users may hit npx cache link errors; the README suggests clearing the cache or pinning a version as workarounds.
• Several promised features—PDF ingestion, folder input, auto-transcription, and continuous updates—are listed as “coming soon (when I have a free night).”

Verdict

Worth a look if you regularly turn long interviews, podcasts, or meeting recordings into reference documentation and want granular control over API spend. Skip it if you need PDF or folder input today; those aren’t implemented yet.