KLR.gg: automated CS2 match analytics for private lobbies

The problem was turning raw server output into ratings people trust, without anyone uploading anything. The data is there: a CS2 server with the MatchZy plugin (2024 docs) writes a log line for every kill, a CSV of per-player stats for every map, and a GOTV demo of every match. But raw numbers are not a rating — kills without round impact, missing round counts, players who change nicknames, several Steam accounts per person. A leaderboard that players do not trust is worse than no leaderboard.

One lobby, one season: HLTV-style Rating 2.0 and 3.0, Round Swing — how much each player's actions moved the round win probability — and a Glicko-2 skill rating that treats every kill as a duel between two players.

It was built as a collection worker, a recompute-from-scratch rating pipeline and a multi-tenant web product. The worker pulls logs, stats and demos from each community server; the pipeline parses logs into duels and computes four ratings — HLTV-style Rating 2.0 and 3.0 (HLTV published the 2.0 formula’s components in 2017), Round Swing, and a Glicko-2 skill rating (Glickman’s system with rating deviation and volatility, 2012). Every run recomputes the whole season, so there is no double counting and re-running is safe.

• Automatic collection. A Python worker connects over FTP/FTPS, detects new files by modification time, pulls them into S3-compatible storage and queues a recompute. Credentials are stored encrypted. Logs (~0.5 MB per map) are the primary source and are kept; demos are the fallback and are deleted after parsing.
• Rating pipeline. pandas modules parse logs into duels, compute round-by-round KAST, Round Swing with an economy weighting for each kill, Rating 2.0 and 3.0, per-map ratings and Glicko-2. About two seconds per map.
• Team balancer. For a chosen set of players it proposes four splits — by Rating 2.0, Rating 3.0, Skill and a mix — with the expected gap between teams. The logic is mirrored in TypeScript so the web does it instantly; shared lineups get a public link with an Open Graph preview.
• Product. Multi-tenant lobbies with roles, Steam OpenID sign-in, statistics snapshots with one-click rollback, notifications, credits paid via Stripe for the render studio (2D round replays), GDPR clickwrap consent with audit, data export and account deletion.
• Steam GC bot. A Node.js bot fetches official matchmaking demos by share code, so ranked games count in the same rating as private ones.

The table compares the spreadsheet most lobbies start with and the automated pipeline KLR replaced it with. A spreadsheet is free and needs no server access, but it depends on one person, cannot compute round-level metrics from a stats CSV, and silently double-counts when a map is entered twice. The pipeline needs FTP credentials and a worker, and then nobody touches it.

From a game server to a leaderboard without a human in the loop: worker, pipeline, API and web are separate services in Docker; the only artifact the web ever sees is one JSON per lobby, served through the API with access checks.

Ingestion is fully automatic and the pipeline keeps up with match nights: a six-map evening is parsed and rated within a minute of the last round. The balancer closes the gap between teams to under one percent of predicted round-win share.

Rating over time per match day, switchable between four metrics, next to the team balancer: the two screens that actually get opened before a match night.

Python 3 rating pipeline (pandas, matplotlib, awpy for demo parsing); worker on ftplib with zstd and MinIO; FastAPI with SQLAlchemy and Alembic on PostgreSQL 17; Next.js 16 with React 19 and Tailwind 4; Node.js Steam GC bot; Steam OpenID, JWT sessions, Stripe. Everything runs in Docker Compose locally and under systemd behind nginx on a VPS in the EU.

This approach is for any team that needs an event-data platform: automatic collection from many remote sources, a parsing pipeline that must be idempotent and fast, scoring models people can trust, and a multi-tenant web product with roles, billing and compliance on top. Strip away the game and the same architecture serves telematics, IoT fleets, trading logs and clickstream analytics. Compliance is part of the design rather than a bolt-on: because player profiles are personal data, account deletion and export were built in from the first release, as the GDPR right to erasure (Article 17, 2016) requires.