End-to-end rebuilds of production ML systems I've worked alongside - churn prediction, demand forecasting, franchise operations. Each project ships with code, a live demo, and a writeup of what worked and what didn't.
Drop a bank CSV into the browser and see where your money goes. Zero server, zero upload - rule-based + in-browser ML categorization (Transformers.js + MiniLM), cross-file dedup, subscription creep detection, month-over-month comparison, user-taught categories. Your transaction data never leaves the tab.
End-to-end churn model for fitness studio members: synthetic data → feature engineering → gradient-boosted classifier → SHAP interpretation. Predicting which members will cancel and why.
Continuously scans problem-oriented subreddits for lab project candidates. Pain-point + TRIZ framing extraction via local Ollama (qwen2.5:7b), bge-m3 semantic dedup, multi-agent validation, daily "paper" sync to kargin-utkin admin for review. Approved ideas graduate to /labs as planned projects - meaning this lab discovers the next labs.
Per-location, per-product hourly demand forecast for a franchise restaurant chain. Time-series ML predicting how many products each location needs prior to the rush.
Document-reading apps that turn a pile of files into an answer, with every figure linked back to the page it came from. Each one runs on its own site.
Upload accident reports, bills, insurance documents and photos. It extracts the key facts, builds a timeline, flags what is missing and links every conclusion to its source. Demo on synthetic data.
claim.kargin-utkin.comUpload bank statements and receipts. It categorizes every transaction, reconciles the accounts, catches errors and links every number to its source document.
bookkeeper.kargin-utkin.comCompare franchise startup costs, fees and rules across brands, with every figure sourced to its exact page in the public disclosure document.
franchise.kargin-utkin.com