Khrystyna Volianska, home
← All case studies

Case study · AI · Litero · 2026

Measure quality, then scale: lifting AI output from C to A, then building a lead channel with no human in the loop

Head of Product (advisory contract), Litero · Aug–Nov 2026

Situation

Litero is an AI writing co-pilot. I was brought in on an advisory contract with two jobs: fix generation quality, and build an acquisition channel that runs on its own.

Complication

Output was graded C/C+: not good enough to pay for when general-purpose chatbots are free. The biggest failure was sourcing: drafts cited sources that didn’t exist or didn’t support the claim, and quoted passages that no source contained. A student can’t hand that in. Quality was argued about rather than measured, so nobody could tell whether a prompt change helped or hurt.

Decision

Measure first, then scale. Scaling content volume before quality was measurable would only have multiplied C-grade output. So I advised on a new generation strategy and put an automated score on every output, which turned quality from a debate into a number every change could be tested against.

What I did

  1. 01

    Made quality a number. 200 essays were graded by professors from US universities, then automatically scored against 200 papers written by the old pipeline, so the jump from C to A is a like-for-like comparison, not a matter of opinion.

  2. 02

    Reset the generation strategy. I proposed a new approach to how the AI writes research; the team rebuilt their engine around it and tuned it against the score.

  3. 03

    Built the references framework. A references-generation framework shipped in the product, benchmarked against Jenni.ai.

  4. 04

    Built my own product on top. A programmatic SEO pipeline on the improved engine, publishing ~150 articles a month with no human input, and a web app to run it, showing costs and output, with a scheduler.

What didn’t work

Prompt-only fixes came first. Rewriting prompts and adding a self-critique pass barely moved the score, because the model was still inventing sources and then defending them. That result is what justified changing the strategy instead of tuning the old one.

Impact

  • Output quality from C/C+ to A/A+ on the automated score.
  • An autonomous SEO channel bringing 15% more leads.
  • ~150 articles a month published with no human input, 90% of them indexed by Google within two weeks.

Learning

In GenAI, the evaluation is the product.

Loading the calendar…