AI-driven candidate screening for a safety-critical enterprise
PhD research, deployed in production.
Problem
Evaluating the candidates for a single vacancy took three to four hours of manual reading.
What I built
A multi-stage LLM pipeline on local models, running entirely inside the customer's security perimeter — no candidate data leaves it. The interesting part was not the generation but the evaluation: deciding, stage by stage, when the pipeline is allowed to be confident.
Outcome
Three to four hours down to under a minute per vacancy, and 89% less manual review across 1,500+ cases.
Cutting a cloud bill by 90% and making the infrastructure reproducible
Early-stage startup.
Problem
An enterprise-shaped AWS setup on a startup's budget, built by hand in the console and impossible to reproduce.
What I built
Removed everything the product did not need, most expensive first, then reverse-engineered what was left into Terraform with AI tooling and LLM-as-a-judge checks on the diffs.
Outcome
90% lower monthly cloud spend, start to finish in two evenings.
An AI assistant with no write access to the source of truth
A small volunteer team tracking inventory in Google Sheets.
Problem
They wanted an AI assistant, but a model with write access to the source of truth was not acceptable.
What I built
An MCP server for Google Sheets: read-only by default, “propose, don't apply” for writes, every change approved by a human before it lands.
Outcome
No path in the design that lets the model corrupt data silently.
Labour-market intelligence pipeline
The data layer underneath my research.
Problem
Questions like “which skills are actually in demand” are unanswerable without a clean, continuously updated corpus of job ads.
What I built
A collector that pulls vacancies from dou.ua and other boards into a knowledge base, then categorises them with managed machine-learning services. This pipeline is what the 2023 and 2025 papers are built on.
Employer matching and career testing at scale
Universum Global — technical architect and lead backend engineer.
Problem
Students pick employers from branding rather than from fit, and employers have no honest signal about who would actually stay.
What I built
An event-driven platform behind a career test that scores 40 factors and matches them against employer profiles built from employee feedback. I owned the backend architecture and the core engineering.