Skip to content
Vrtron

In development · AI · Engine

NUTS

A next-generation AI poker engine. Real-time, exploitative and self-improving, designed to drop into established poker infrastructure and run live at scale.

year
2026–
role
Concept, research, engineering
scope
Engine · Operator infrastructure · Consumer products

01 · Overview

NUTS is a poker decision engine being built on the strongest ideas from a decade of academic poker AI research, sharpened for the workloads existing commercial tools handle badly: real-time play, exploitative adjustment, and multi-table operation at scale.

The field today splits into offline solvers that compute beautiful equilibrium strategies but cannot play, and academic milestones that proved superhuman play possible but were one-off research artifacts. NUTS is being built as production software that picks up where those papers left off and ships as something operators and product builders can actually deploy.

Designed from day one for 6-max, not as a heads-up special case retrofitted later. Designed to slot into established poker infrastructure rather than demand new infrastructure to host it. Designed to keep getting better while it runs, not just at training time.

02 · Highlights

Features in development

01

Real-time, not precomputed

Continual depth-limited re-solving in the modern lineage: a guiding blueprint plus a neural value head plus live search. Coverage and specificity together, in the time a human takes to blink.

02

Sub-millisecond decisions

Targeting typical decision time per node well under existing real-time solvers, including opponent adjustment. The difference between an engine that can drive a live game and one that cannot.

03

Exploitative by default

An integrated opponent loop observes every action, blends against a Bayesian prior, and tightens villain ranges and sizing models live. Against unknown opponents it defaults to balanced equilibrium; against players with leaks it drifts toward the best response.

04

Self-improving arena

Built-in arena pits the engine against snapshots of itself, against external bots, and against captured hand histories. Every session generates training data. Multi-stage, auto-chained training: the engine improves overnight without manual intervention.

05

6-max native

Heads-up is the historic benchmark; 6-max is what people actually play. NUTS is being built with 6-max as the primary target across range conventions, opponent model, and arena, with strict fairness boundaries between bot instances.

06

Drop-in shape

Existing site adapters and poker tooling speak well-known protocols. NUTS's decision API is being designed to match the standard wire shape, so existing integrations can swap engines without code changes.

07

Strict fairness

When multiple NUTS seats share a table, each instance only sees what that seat would see. No shared hole cards, no shared future plans, enforced at the API layer. Accidental collusion is impossible to build by accident.

08

Observability and audit

Decision logger, regression detector, continuous benchmark, shadow harness. Reproduce results, run regressions, push changes into production with audit trails rather than blind upgrades.

03 · Stack

Made with