How We Built Our Best Pokémon AI: The Trainer's Edition

“The best trainers do not just teach their Pokémon how to attack. They learn how to listen to the battle.”

Wuh gine on! Anaiya here. 👋🏾

A few weeks ago, we shared our technical recap of the Pokémon TCG AI competition on Kaggle in When Simple Beats Clever. But if you love Pokémon or play the games, reading about tree searches and gradient-boosted rankers does not capture what it actually felt like to fight in that stadium.

So here is the full story of our run, told strictly in the language of the games: how we trained an adaptable Eevee mind, the embarrassing Thunderbolt blunder we had to patch, and why an unglamorous rock-bug became our ultimate MVP.

1. The Eevee Strategy (Starting Without a Locked Type)

Many baseline rule engines in competitive card play lock their heuristics to a single deck archetype. If you give them a Charizard deck, they know the specific sequencing. But if you switch them to a Water or Electric team, rigid card-specific rules quickly break down.

We decided to train our battle program to think like an Eevee.

In the games, Eevee does not lock itself into being a Water, Fire, or Electric type from birth. It waits and watches. When it touches a Water Stone, it adapts into Vaporeon. When it touches a Thunder Stone, it becomes Jolteon.

Our battle engine did the exact same thing with any 60-card deck you handed it. It looked at the cards, checked who the primary evolution line was, figured out whether the team wanted to rush the opponent with fast attacks or build up huge power on the bench, and derived its entire battle posture on the fly. It adapted its strategy like an Eevee choosing its evolution stone in the middle of a battle.

2. The Overcharged Pikachu Blunder (Our Biggest Rookie Mistake)

Even the best trainers make goofy mistakes early on. When we reviewed our match replays from the stadium floor, we found our trainer making an embarrassing blunder.

Picture this: your active attacker already has three Energy cards attached, which is all it needs to unleash its strongest knockout attack. But on the very next turn, the trainer shouts: “Attach another Energy to the active Pokémon!”

It was like standing in a gym battle screaming at Pikachu to absorb more electricity when Pikachu’s cheeks were already crackling with maximum voltage! Meanwhile, your backup Pokémon on the bench had zero energy attached, leaving them completely unable to evolve or fight.

To make things worse, this particular deck dealt extra damage based on how many cards you held in your hand. Every time our trainer threw away an Energy card on a Pokémon that did not need it, our attack power actually dropped!

As soon as we spotted this in the match tapes, we reprogrammed the trainer’s battle instinct. From that day on, whenever the main attacker had enough power to strike, every extra Energy card went straight to charging up the benched reserves.

3. Professor Oak’s Lab vs. Ash Ketchum’s Field Instinct

Back in the laboratory, we built a computer simulator to run practice battles against training dummies.

One day, our lab computer showed that a fancy new search tactic gave us a giant +10% boost in mock battles. We thought we had cracked the code! We were so confident that we entered that exact build into the live Battle Tower that weekend.

It got utterly destroyed. In fact, it posted our lowest tournament score of the entire season.

Why did that happen? Because practice dummies in Professor Oak’s quiet lab do not fight like real trainers in a packed stadium. The simulator math was too clean and made our confidence look twice as wide as it really was.

Like Ash Ketchum, we learned that classroom chalkboards can fool you. Ash never won championship badges by staring at damage charts. He won by heading out into the stadium dirt, taking tough losses, watching match tapes from the sidelines, and listening to how real battles actually flow. From that point on, we never trusted a lab test until we proved it against live opponents on the real tournament ladder.

4. The Pikachu vs. Crustle Paradox (Our True MVP)

Every Pokémon fan wants their flashy Pikachu moment. We fell completely in love with our high-tech psychic ace: a Stage-2 Alakazam team that held bending spoons and scaled attack power off huge hand sizes. We spent months thinking Alakazam was our crown jewel.

Then we looked at the real tournament data across hundreds of matches, and the numbers told an unexpected story:

Our top internal performer was Crustle.

Crustle is a humble Bug and Rock type carrying a giant slab of stone on its back. It is not glamorous, and nobody puts it on the front of the game box. But on the live ladder, Crustle was an absolute bulldozer. It did not need tricky five-card combos or delicate setups. It just stepped onto the battlefield, put up its heavy rock armor, and started taking prize cards turn after turn.

In fact, the data showed that even when Crustle had an unlucky streak, its worst score was still higher than Alakazam’s average!

5. Midnight at the Championship Gate (The Final Two Pokéballs)

When the tournament reached its final deadline, the platform enforced a strict rule: only the final two submissions registered on your trainer belt at midnight would battle on the post-deadline evaluation ladder. After that, the gates locked, and those two teams had to battle for two full weeks to establish final placement.

We strapped two Pokéballs to our belt:

1. Pokéball One (The Tank): Our rock-solid Crustle tempo team, sent in to lock down our rating ceiling.
2. Pokéball Two (The Tactician): Our Stage-2 Alakazam combo team, sent in to test whether our trainer could pilot complex, high-IQ psychic setups without panicking.

Just hours before the deadline, we had tested a last-minute tweak, but the live ladder showed it was losing points. Instead of stubbornly hoping for the best, we swallowed our pride and rolled back to our trusted build. That single rollback saved our team over 37 rating points when the final scores settled!

6. The Final Tournament Score

When the two-week evaluation period ended, the final ratings on our submission tracking logs confirmed the results:

• Crustle (v104) settled at a peak score of 615.8, giving our team its top standing across our submitted models.
• Alakazam (v109) held its ground through complex matches and settled at 583.4.
• Our rollback decision was verified by a margin of 37.6 points over the discarded experimental candidate.

In the end, competing effectively on the Pokémon ladder was not about finding an unfair trick. It was about training an adaptable Eevee mind, fixing rookie mistakes, respecting the quiet power of a sturdy Crustle, and trusting real battle experience when the stadium lights turn on!

Catch wunna on the ladder, keep your battle instincts sharp, and walk good! 🌊

Warmly,
Anaiya ✨