Vibechess AI tutor for chess
Log inStart free
VibeChess Lab

Chess intelligence, from the board up.

We're building the layer between chess engines and language models: structured chess knowledge, position-aware context, evaluation benchmarks, and chess-specialized models.

Join the lab
A repeating pattern of matte coral and mint pawns on a cobalt background
One player's chess, as a linked graph Sessions and openings link to the concepts you've trained; coral edges show what you've mastered.MASTEREDMASTEREDMASTEREDMASTEREDLEARNINGLEARNINGLEARNINGLEARNINGLEARNINGLEARNINGNEEDS WORKYouNajdorfGameReviewRuy LopezGameReviewPuzzleStreak ×12Mate-in-2PuzzlesItalianOpeningDrillCaro-KannDrillFrenchDefenseStudyK+PEndgamePracticeBlitz vsMaia 1700SicilianDefenseRuy LopezItalianGameCaro-KannFrenchDefenseQueen'sGambitDeclinedKnightForkPinSkewerBack-RankMateGreek GiftSacrificeDiscoveredAttackKingSafetyKnightOutpostBishopPairSpaceAdvantageProphylaxisIsolatedQueen'sPawnMinorityAttackPassedPawnOppositionYouSessionOpeningConcept
Fig. 01 — One learner linked to their sessions and concepts; edge weight is mastery.
LAYER 01 / SUBSTRATE

Chess Graph

A structured knowledge graph for chess competence.

  • Opening repertoire, strategic understanding, recurring patterns
  • Interlinked, personalized, and retrievable
  • Queryable substrate the entire stack reads from
Tech
schemaontologyknowledge graphgraph retrievalgraph analyticsrerankingANNHNSW
Every piece, scored Cell colour and value show each piece's contribution; mint favours White, blue Black.abcdefgh12345678-5.02-5.18-6.89-4.10-1.55-2.03-1.23-0.55-0.63-3.99-3.90-1.61-0.81+1.56+1.11+3.51+0.76+0.71+2.83+0.62+0.95+3.63+6.25+0.60+3.95+4.21
Fig. 02 — Every piece scored by its contribution to the position.
LAYER 02 / RETRIEVAL

Chess Context Assembly

Board-state retrieval and analysis.

  • Designed for LLM reasoning and explanation
  • Surfaces exactly what a model needs from the position
  • Not generic RAG pointed at a position
Tech
Stockfish (NNUE)Lc0Maiaboard-feature extraction
Where the engine spends attention Per-square activity over a middlegame, summed across the search.ABCDEFGH12345678 hi lo
Fig. 03 — Per-square engine attention over a middlegame.
LAYER 03 / MEASUREMENT

Chess Evals

Benchmarks for chess coaching in language models.

  • Scored move by move against expert annotations
  • Position-grounded scoring to catch hallucinations
  • Level-calibrated, from beginners to titled players
Tech
eval harnessscoring / judgebenchmarkssamplingobservability
Reading a real board A simple rook endgame, detected with class and confidence — the way the camera sees it.wK 0.99wP 0.97wR 0.95bK 0.98bR 0.94model vibechess-vision · 5 detections · 99.2% mAP
Fig. 04 — A position read from an image, each piece boxed with class and confidence.
LAYER 04 / MODELS

Chess Models

A family of chess-specialized models.

  • Chess-language adapters
  • Personalized feedback models
  • Vision models that read positions from images
Tech
LoRA / QLoRAfine-tuningLc0YOLO
PRODUCTS

Three products, one stack.

The research above becomes products — the foundation model, the knowledge layer, and the agent players talk to, all built on the same stack.

Foundation

VibeChess Core

A chess-native language model that reads any position and explains the why — not just the best move.

A matte mint rook on a cream plinth, coral background
Knowledge

VibeChess Graph

Turns a player's games into a personalized knowledge graph: strengths, gaps, and what to study next.

A matte coral queen on a charcoal plinth, cream background
Coach

VibeChess Agent

An agent that teaches — it plays, reviews your games, and answers the follow-up questions.

A matte mint king on a cream plinth, cobalt background
Collab

Join the lab.

We’re open to connecting with experts in knowledge graphs, AI agents, machine learning and chess.

Get in touch
Directanna@vibechess.meLinkedIn ↗

Get in touch

Drop us a line — we read every message.

No spam — we reply personally.