Ruslan Kovychev · backend and ML engineer · kovyrus@gmail.com · LinkedIn
Write-ups of three commercial projects, each described the way I would explain it to another engineer: what the problem actually was, what I chose, what went wrong, and what I would change. Where I owned only part of a system, that is stated at the top of the write-up.
Client names, client code and client data are not published here. What follows is architecture, reasoning and the parts of the work that are mine to show.
2026 · sole architect and developer · Python, FastAPI, PostgreSQL, React, Docker
A platform that tracks competitor video content across two social networks, stores engagement as a time series so that growth rather than absolute size drives the ranking, and alerts when a video breaks out of its expected curve. Includes the detector that had to be rebuilt three times, and why the first two versions were confidently wrong.
2025 · subcontractor on one slice of an agency platform · Python, spaCy, pymorphy3, SBERT
Telling an SEO team not just which words they are missing but where on the page, by labelling text with a structural role and weighting each role. Includes the measured answer to whether the language model in the pipeline earned its place, and the finding that 95 percent of runtime sat in one unbatched loop.
2025 · PyTorch, PyTorch Geometric
Forecasting the motion of many interacting articulated objects, where the interaction graph is rebuilt from the model's own predictions at every step, so error changes the topology and not just the coordinates. Trained on 10-step rollouts, evaluated on 20.
RLS Inspector · a static analyzer that reads Supabase migration files and reports Row Level Security defects with severity, evidence and a fix. Six deterministic rules, no language model in the decision path, tuned against a fixture of known defects.