lasteverLM
Material in. Understanding out. PDFs, text, web pages, images and audio go in; questions, reports, quizzes, tables, mind maps, infographics and spoken overviews come out, each sentence traceable to the passage it came from. A NotebookLM-style assistant, built to be checked: every claim carries its passages, every document its generation log.
- Live: lasteverlm.zehentleitner.co — login by magic link on an allowed domain.
- Code: github.com/oliver-zehentleitner/lasteverLM — one repository, six services.
- Docs: oliver-zehentleitner.github.io/lasteverLM — this site: the project in one page (features, principles, security, how it was built), architecture and API contract, the RAG pipeline, user interface, walkthrough script, repository standard.
- Why: the Keep the Why dashboard — every decision with the alternatives it rejected, in the order it was taken.
How it is built
Everything runs on one Vultr account: a Kubernetes cluster, Managed PostgreSQL with pgvector, Object Storage and Serverless Inference for the language model, the embeddings and text-to-speech. Delivery is containers only: every service is an image in GitHub's registry, deployed with one Helm chart, released with one tag. Every process scales horizontally; PostgreSQL is the only state and also the job queue for the workers.
| Directory | Purpose |
|---|---|
core/ |
shared library: models, settings, auth, rate limiting, usage accounting, provider interfaces, object storage |
account/ |
magic-link login, sponsored credit, admin panel backend, access log, blocks and bans |
infra/ |
Vultr cluster, Helm chart, deploy automation, migrations runner, smoke test, local compose stack |
ingest/ |
upload, URL and image sources, parsing, chunking, embeddings, source summaries, ingest worker |
rag/ |
retrieval, cited answers, generated documents, audio, sharing, rag worker |
web/ |
wizard, two-column notebook view, viewer, login, credits, admin panel |
The contract between the services is Architecture and API contract; how a question becomes a cited answer is The RAG pipeline; the screens are in User interface; what every directory must carry is the Repository standard. The walkthrough script is the ten-minute recorded tour.
How it was built
Seven coding-agent sessions working in parallel, one per directory plus
a head that reviews and merges, with a human maintainer who decides and
tests. Each directory keeps its own context/ with the reasoning behind
its code; the dashboard above reads all seven.