Iurii Uses: #AI

Our support inbox was drowning — most tickets were repeat questions our docs already answered, but customers weren't finding them. Iurii built us a retrieval-based assistant on top of our existing knowledge base in about three weeks. It answers around 60% of incoming questions correctly and hands the rest to a human with full context already attached. What I appreciated most was that he was honest about where LLMs would fail before we started, so there were no surprises later.

We wanted to add an AI feature that turns messy user notes into structured records, but our first attempt returned unpredictable JSON that broke the app half the time. Iurii rebuilt it using structured outputs against the OpenAI API with proper validation, so the data is always shaped the way our database expects. He also added a fallback path for when the model is unsure instead of letting it guess. It's been running in production for a month with no manual cleanup.

I'd read about llms.txt and AI discoverability but had no idea whether any of it actually mattered for our SaaS docs site. Iurii's audit gave me the honest answer: most of it is speculative right now, but a few things genuinely move the needle today — clean semantic HTML, robots.txt directives for AI crawlers, and structured data that retrieval systems can use. He also flagged that we were blocking GPTBot by accident in a leftover robots rule from 2023. The report saved me from chasing trendy fixes and pointed at the boring ones that actually work.

We process thousands of supplier invoices a month and the team was keying them in by hand. Iurii built a pipeline that reads each document, extracts the fields we care about, and routes anything that doesn't match a known template to a human — with a confidence threshold so nothing gets posted automatically that the model wasn't sure about. He was blunt up front that we shouldn't try to automate the low-confidence cases, and that honesty is exactly why finance trusts the system. It's handled around 85% of the volume for two months now with no bad postings we've had to reverse.

I'd built most of our MVP with Cursor and it looked finished — it compiled, the tests were green, the demo worked. It just wouldn't survive real users. Iurii took it over and found what I couldn't see: authentication handled four different ways, tests that only asserted what the code already did, and a dependency list half of which was unused. He didn't rewrite it from scratch — he told me honestly what was salvageable, ripped out the dead code, and got it to something a real team could build on. Two weeks and it went from 'looks done' to actually shippable.