Projekt-Analyse

Lektobot

An AI proofreading agent for Slovene academic Word documents — two-pass review, a full audit trail, and output that is provably identical to the original except for the comments

Übersicht

Lektobot automates a real expert workflow: proofreading Slovene academic texts. It reads a Word document, finds likely language issues with a two-pass AI review — one model proposes, an independent model audits every proposal — and writes an annotated copy containing nothing but Word comments. It never edits the text, and it refuses to release output it cannot prove is unchanged. On a real 31,204-word master's thesis it inserted 250 reviewed comments in 1 hour 24 minutes, with every suggestion traceable from proposal through audit to the final comment — work that would otherwise take a proofreader days of repetitive first-pass reading.

Projektdetails

Client
Niro Digital Product
Industry
Language Services / Document Automation
Service
AI Workflow Automation
Year
2026
Services
AI Workflow AutomationAI Agent DevelopmentDocument Processing Automation
Lektobot CLI — completed review of a 31,204-word thesis with 250 inserted comments and all 11 integrity checks passing

A real run: 31,204 words reviewed, 250 comments released through the validation funnel, integrity PASS on all 11 checks

An automation you can trust is one that can prove what it did — and prove what it didn't touch.

Lektobot review output with the candidate funnel — 400 proposals filtered down to 250 released comments

Two AI Passes, One Deterministic Judge

One model role proposes corrections over the document; a second, independent model audits every proposal against the source and accepts, revises, or rejects it with a reason code. Deterministic validation then re-anchors every suggestion to the exact characters it refers to, enforces thresholds, and removes duplicates. Nothing is taken on faith: on the thesis run, 400 proposals entered the funnel and 250 verified comments came out — and every drop is recorded with a reason, so the funnel always adds up.

Engineering the Trust

Preservation

Provably Unchanged Documents

Before any output is released, 11 independent checks must pass — including byte-level inversion and a per-character formatting ledger — proving the visible document is identical to the original except for the comments. If one check disagrees, nothing ships.

Cost

Price Known Before It Runs

An inspect command estimates the cost of a review before a single model call leaves the machine, and a hard per-job spend ceiling is enforced against an append-only call log that cannot drift.

Resilience

Interruptible and Resumable

Every model result is checkpointed the moment it arrives. A rate limit, an outage, or Ctrl-C leaves a job that resumes exactly where it stopped — completed segments replay from disk in milliseconds.

Security

Documents Are Data, Not Instructions

The manuscript is delimited as untrusted content in every prompt: instructions found inside a document are proofread, never followed. Malformed and malicious files are rejected before anything leaves the machine.

Privacy

An Audit Trail Without the Text

Every run writes a full audit report — hashes, versions, the candidate funnel, per-role costs — containing no manuscript text at all. A runtime guard rejects content-shaped values, and a test asserts it.

Quality

Versioned Review Profiles

Five review profiles — from conservative academic to thorough — are versioned configuration: categories, thresholds, density caps, and pinned model IDs. Changing a model is a config edit, not a code change.

Lektobot inspect, jobs, and verify commands — cost estimate before running, resumable jobs, independent verification

Built Like an Operations Tool, Not a Demo

The workflow is fully operational: estimate the cost before committing, run with bounded concurrency, resume interrupted jobs, and independently verify any output file against its source — even years later, without the original job data. Model failover is defined per role with pinned fallbacks, so an outage changes which model answers, never what it is asked to do.

What This Shows About Our AI Automation

Lektobot is what we mean when we say we automate workflows with AI. We took an expert's repetitive first pass — reading tens of thousands of words hunting for comma errors — and gave it to supervised models, while keeping the human expert as the final authority over every suggestion. The same pattern applies to any document-heavy process: contracts, reports, applications, translations. AI does the exhaustive pass; deterministic checks prove nothing was corrupted; people make the calls.

Have a document workflow that eats expert hours?

Review, intake, compliance checks, summarization — if your team reads piles of documents to find what matters, that first pass can likely be automated with AI, verifiably.

Scroll handle
0