VIVEKJOSHY

Open-source Bayesian rating library for asymmetric multi-team, multiplayer matches, implementing the Weng–Lin models with extensions. Published in the Journal of Open Source Software; preprint on arXiv since January 2024.

Independently ported to 7 languages by other people — TypeScript, Luau, Go, Kotlin, Elixir, Java, and C#. Cited 11 times, including in IEEE Transactions on Games, COLING, and the Journal of the Royal Statistical Society. ~114k PyPI downloads a month.

Pinned as a dependency in Ray/RLlib and used as the rating layer in Neural MMO; both are checkable in those projects' dependency manifests. Reported in production use at OpenAI, MultiVersus, Hunt: Showdown, and BeyondAllReason, and the rating engine behind the Vairified platform.

openskill.me  •  github.com/vivekjoshy/openskill.py  •  DOI: 10.21105/joss.05901  •  arXiv:2401.05451

• Own the backend ML platform: shipped the Partner API to production behind OAuth and scoped API keys, with maintained Python and TypeScript SDKs.
• Built VAIR, the production rating engine: Thurstone–Mosteller via OpenSkill, with score-margin weighting and adaptive uncertainty.
• Built Sentinel: extracts implicit business logic from three codebases by AST analysis and renders it in natural language for non-engineer review — the interpretability problem in miniature, on a real system.
• Led the legacy PHP migration off a 184-file codebase.
• Delivered multisport sport-filtering end to end, from schema through cross-sport search to per-sport rating buckets.

• Created matchmaking algorithms and skill assessment models.
• Built prototype system using Python and PostgreSQL.

• Maintain OpenSkill, a peer-reviewed Bayesian rating library with production deployments across games and sports.
• Contributed role icon support to hikari, a Python Discord API library (merged, 2021).

• Silver medal: 61st of 1,950 teams.
• Implemented domain adversarial training with gradient reversal.
• Designed a custom molecular data processing architecture.

• Built CDCL-based reinforcement learning system for SAT/SMT problems.
• Custom rewards created for efficient solution space exploration.

• Built an entropy-based framework for token evidence aggregation.
• Created logarithmic pooling for confidence scoring.

Owner and administrator of discord.gg/neuroscience, the largest theoretical neuroscience community on Discord, with over 6,940 members.

Expert role on the MLST community server.

Contributed to the Common Gotchas chapter (docs.python-guide.org)