Skip to content

The project

lasteverLM turns material into understanding you can check. PDFs, text, web pages, images and audio go in; questions, reports, quizzes, tables, mind maps, infographics, spoken overviews and decisions come out, and every sentence can be traced to the passage it came from. This page is the whole picture in one place: what it does, how it is built, the principles behind it, how it was made, and what it deliberately leaves out. Details live in the linked pages; the reasoning behind every decision is in the Keep the Why dashboard.

What it does

Sources

  • Five kinds: PDF, text, web page, image, audio. A URL may also return structured data (JSON, NDJSON, XML with RSS and Atom, CSV, TSV, YAML): every value is read on its own line with its key path or its column names, so a cited number never stands without its context; hostile input (entity expansion, alias bombs, deep nesting) fails with its reason.
  • PDF capture: the text layer, the vision model for scanned pages, or auto (vision only for pages without a legible text layer; a layer of glyph codes counts as not legible).
  • Images are transcribed and described by the vision model. A large image or a dense scanned page is read in overlapping tiles, because the platform shrinks every image to a fixed budget and small text would otherwise be invented rather than read (measured in the vision evaluation). Every model-made text says which model made it and that it is a description, not a verbatim transcript. A transcript that loops into repetition is retried once and then fails loudly; it is never indexed.
  • Audio is transcribed by a speech-to-text model. MP3 and WAV go in as they are; M4A, AAC, OGG, Opus, FLAC, WebM, the audio track of MP4, and WAV over the platform's 25 MB cap are converted to MP3 first by a minimal ffmpeg. The converted file and the original stay available to play and download, so the transcribed audio can be checked.
  • Library per user: every source is stored once (deduplicated by content or normalised URL), linked into any number of notebooks, deleted into a trash for a user-defined number of days. Nothing is shared between users.
  • Read when used: adding a source stores it and costs nothing. It is read (extracted, chunked, embedded and charged) only when a question or an output needs it, after the price has been shown. Without credit it waits and is read as soon as credit is there.
  • Web pages have versions: a page is fetched when it is first used. "Re-read if changed" fetches it again: unchanged, nothing is charged; changed, it becomes a new version that takes over the page's links, while the old version stays readable. Every read version keeps an inert snapshot of what was read (the fetched HTML rebuilt from an allowlist: text, structure and links, no scripts, styles, frames or images), so the original a citation points to survives the live page changing or disappearing.
  • Starter notebook: every new user gets their own copy of a notebook about lasteverLM itself: the complete technical handbook as a PDF, an FAQ, the public price list as a web page, a screenshot of the app, and finished outputs of every kind, ready to ask and to check.
  • First steps: on the first sign-in a three-step introduction outlines where sources go, where the format is picked and where Create is. The help page (the ? at the top right, next to the credit chip and the profile menu) has a short guide, shows the introduction again and recreates the starter notebook if it was deleted.

Outputs

  • Kinds: report, quiz, flashcards, data table, chart, mind map, infographic, audio, decision.
  • Options: output language, for audio the style (overview, story, conversation with two voices, FAQ with one voice asking and one answering), voice and speed; for reports style and length; for quizzes number of questions, difficulty and audience; for every kind free notes that steer focus and tone but never the citation rules.
  • Decision: a question weighed against the sources. The platform's decision model gives the number (a probability, a choice with its distribution, or a level on a scale); the language model writes the evidence for and against independently from the same passages; both go through the support check, and their agreement is shown (consistent, conflicting, unclear, or "the sources do not decide this"). Without supported evidence a decision carries no number; with only one supported item the number is shown as a tendency.
  • Chart: "Find charts" proposes number series from the sources as small previews; the user picks which are drawn. Every value is checked exactly against its passage (in any notation: 1.906, 1906, 1,9 Mio., percent); a computed value is marked as computed, never as supported.
  • Verification levels: every user sets how much checking matters and what it may cost; each output and chat answer can change it, and each level shows its price before Create.
  • Basic (free): the word check plus a negation check that parses each sentence (spaCy, English and German) and marks a paragraph whose negation contradicts its passage as a possible contradiction (negation evaluation).
  • Standard: every paragraph the deterministic checks doubt (unsupported, other language, possible contradiction) is judged once by a model of another family than the author.
  • Thorough: every paragraph is judged; failing ones are revised and judged again, at most three rounds, and what still fails is removed and listed as removed.
  • Bulletproof: two judges of different model families judge every paragraph in parallel; a paragraph counts as supported only if both agree, and a disagreement that survives three rounds is kept and marked disputed, with both verdicts shown. The judge models were chosen by published evaluations (judge, second judge). A paragraph the judge does not find in its cited passages is searched for in all selected sources and judged again; if it is found there, it counts as supported with the citation added and marked, and only what is in none of the sources stays unsupported. A contradiction is never searched away: a conflict between sources stays visible. With a judge verdict only the judge's badge shows; the word check never claims meaning. "Check with a judge" on an existing output or chat answer annotates only, runs as a background job and shows its progress.
  • Coverage: every output shows how many of the selected sources it used and names the unused ones with the generator's reason, so no source is left out silently.
  • Data tables and infographics first propose candidates from the selected sources, like charts; the user ticks the tables or themes to build, each from its own sources.
  • Every output has an ID, shown in its head and after every error, copied with one click; every failed attempt is stored with it, so support finds what happened after any deploy.
  • Outputs never feed new outputs. A generated document holds no knowledge its sources do not, so outputs are made from original sources only; the chat can still talk about an output.
  • Voices: the defaults per output language come from a listening test (German Dylan and Ono Anna, English Ryan and Vivian), with the recommended voices listed first; a user's own choice always wins.
  • Outputs belong to the user, live in the library, survive the deletion of the notebook they were made in, and can be added to other notebooks as sources.
  • Sharing: a single output or a whole notebook as a read-only link, readable without login, with all its evidence.

Chat

Conversations per notebook as tabs, scoped to the sources the user selects, with numbered citations into verbatim passages, the same support check as the outputs, copy, thumbs up and down, and three suggested follow-up questions. The chat can also act on its notebook: asked to add a web page, take entries from the library, create an output and explain it, it answers with a plan (steps, estimated prices, invalid steps struck through with the reason) and runs it only after the user confirms; tool arguments come only from the user's message and library, never from source text.

Viewer

Sources and outputs the user opens appear as tabs. A source tab shows the original (PDF inline at the cited page, image, audio, a text or Markdown file inline, a web page as the snapshot of the version that was read, in a sandboxed frame without scripts, with a link to the live page) and the text as the system read it, chunk by chunk; a citation highlights its chunk.

Administration

Users and roles (user, viewer with a read-only admin panel, admin), who may sign in and who starts as viewer (domains and single addresses for both), a finance area (below), rate limits, blocks and bans with their rules editable at runtime, sign-in attempts, visits, mail configuration, and a System page with the installed version, every pod and a live architecture diagram that colours each component by its state.

Finance

Every booking keeps the measured cost in US cents (the platform's currency), the EUR/USD rate applied, the cost in euro and the price charged to the user, all fixed at booking time, so a later change of rules never rewrites history. The charge is the cost plus a margin, set globally and per action (a chat answer, each output kind, each source kind, the judge, a decision), optionally with a fixed fee per action. Users see the charged price before every paid action, a public price list, and their own Credits page (balance, voucher redemption, spending by day, by action and entry by entry). The admin sees costs, prices with a live preview, revenue against paid and free vouchers, and profit after the platform's infrastructure costs, which are read daily from Vultr's billing API. Voucher codes are encrypted at rest with a separate key, so an admin can show them again and send them by mail. Adding a source is free; reading it is charged when it is first used.

Principles

Everything derived from a user's input can be checked by that user. Every unit of every output (a report paragraph, a table row, a mind-map node, a sentence of an audio transcript, a piece of evidence in a decision) carries the passages it rests on and a support status from a deterministic check: supported, unsupported, or unverified when the output was written in another language than the material. Units that fail are marked, never dropped. Every output carries its generation log: retrieval candidates and which entered the prompt, the model, the prompt version, usage, timing and the check's counts. The shared page shows all of it.

Fail loud. A wrong or missing value stops the process instead of falling back silently: a service without a price for its model does not start, a loader that extracts no text fails the source, a provider answer without usage is an error.

The cheaper path by default. Whatever costs extra is an action the user takes, never a default: sources are read only when used, a web page is fetched again only on request, the verification level starts at the free basic level, and every paid action shows its price first. An action starts only when the credit covers its estimated maximum, and then runs to its end, even below zero, so nothing stops half-paid.

Measure, do not assume. Five billing facts were corrected by a measurement instead of the documentation: text-to-speech is charged per second of speech, not per character; the decision model reports input tokens, although the catalog says completion; speech-to-text is charged per second; the main model's input price differs from its output price; the platform prices in US dollars while credits are in euro. Each is recorded with the measurement. The judge's models were chosen by an evaluation rather than by reputation: on 24 hand-written claims the word check alone caught none of the negated, swapped or off-by-one claims; no judge level missed a negation, and two of the three scored every claim correctly (details and limits in the evaluation).

One platform. Kubernetes, PostgreSQL with pgvector, object storage and every model run on Vultr. Web search was considered and left out because it would add a second vendor.

Architecture in short

Six parts in one repository, each with its own context: core (the shared library), account (sign-in, credit, administration), ingest (sources, library, extraction, chunks, embeddings), rag (retrieval, chat, outputs, sharing), web (React), infra (OpenTofu for the platform, Helm for the application, the pipeline). PostgreSQL is the database, the vector store and the job queue. A Git tag builds and deploys; a smoke test over the public domain closes every deploy. Details: Architecture and API contract, The RAG pipeline, User interface.

Security

TLS with HSTS and an enforced Content-Security-Policy with hashed inline scripts; magic-link sign-in with one-day sessions and an allowlist; roles checked in every service; rate limits, blocks and bans per address (IPv6 by /64) behind the PROXY protocol, so a client cannot choose its address; SSRF protection on URL sources; HTML from models sanitised before rendering; pods without root, with read-only file systems and NetworkPolicies; secrets in Kubernetes Secrets only; voucher codes encrypted at rest with a separate key; e-mail addresses, IP addresses and answer texts masked for read-only viewers at the server; no paid model call without a configured exchange rate, so nothing is paid upstream and left unbooked; actions pinned by commit with Dependabot, images with provenance and SBOM; code scanning at zero open alerts.

Reliability

Every job heals itself: a running job keeps its claim with a heartbeat, only the worker holding it may store the result, a rollout gives workers time to finish and hands a job over cleanly when it cannot, and a worker that takes over resumes from a journal of the model calls already paid, so nothing is paid twice. Details in the architecture.

How it was built

One maintainer and seven coding-agent sessions over three days: one agent coordinating, reviewing and merging, one per part of the system. Every change went through a pull request that was merged only when every check had passed; every decision was written down with the alternatives it rejected, and is browsable in the Keep the Why dashboard. More than 640 pull requests and 80 tags; a fix is typically live four minutes after its tag. The walkthrough is the script for a guided tour. The full account, with numbers, is in How it was built.

Deliberately not built

  • Web search as a way to add sources: a second vendor and search quality outside the system's control; the plug-in point is described in the architecture.
  • Approximate nearest-neighbour index, reranker, model tokenizer for chunking: named in the RAG pipeline with the size at which they would pay off.
  • Node autoscaling waits for the database's private network attachment.