Poker AI lab. Coach product.
Opponent modeling from 6.6 million analyzed decisions — wired into exploit reads, leak drills, and pressure-test sim. Not a spot pack. Not raw solver output.
The exploit models. Both of them.
Two adapter sets trained on 6.6 million analyzed decisions. Not a paper, not a demo — the trained weights, for sale.
Qwen3-4B per-street swarm
Four LoRA adapters — preflop, flop, turn, river — on Qwen3-4B-Instruct-2507. Rank 32, PEFT safetensors, ~126 MiB each. The deployed set that plays the sim table.
79.7% strict exact-match on the full 11,000-row PokerBench test (preflop 89.3 · flop 83.6 · turn 79.8 · river 77.4).
Qwen3-8B combined
One LoRA adapter on Qwen3-8B, rank 32, trained across all streets in a single model. PEFT safetensors (best_em checkpoint).
The clean single-model run: same street coverage, no per-street routing.
Adapters only — bases are the open Qwen3 weights. Delivery and licensing terms on contact.
// contact@pokagto.coach
Exploit reads grounded in opponent modeling.
We analyzed how 36 frontier models play poker and built archetype models — the coach tells you how to attack stations, nits, and regs in plain English, with your measured leaks when you have data.
- +Archetype exploit reads — not generic ChatGPT
- +6,000+ curated strategy sources behind every answer
- +Personalized leak reads with real accuracy % and sample sizes
- +Always-on dock + full chat at /ask
> ask
How do I beat a calling station?
> coach
Value bet thinner and bigger — they pay off dominated pairs and draws. Cut bluffs unless you hold strong blockers; stations rarely fold. Size for extraction, not balance.
See it on your own game.
$79 a month, full access. Your leak profile starts building from drill one.
Start training →