Inspiration
We've both played with bots in shooters, and they've always been kind of useless as teammates. For me (Steven), CS:GO was an integral part of my childhood, and the bots always struck me as useless. For context, bots have been around since Quake in 1996, and thirty years later you still can't really talk to them. In CS you get a radio menu, and half the time they just decide to ignore it anyway. Even PUBG's new AI teammate only had 18.5% of players say it felt good.
And comms are pretty much everything. Studies keep finding that better teams talk more and faster. Riot is already researching bots that move and fight like people, so we wanted to build the piece that's missing. A robot who does and listens.
What it does
RelAI is a voice teammate you play alongside in a top-down 2D version of Bind's A site. You hold Space and talk to it like you would a friend on comms.
Tell it "smoke Heaven and Truck" and it does exactly that. If you say something vague like "smoke over there," it asks "Smoke where?" instead of guessing. Call the wrong site and it tells you "That's B. We're on A." It also calls out enemies, but only ones it can actually see.
The fun part is that you can talk to an AI! On the same round, if I don't say anything, the team walks into held angles and gets wiped. Make one good smoke call and you win.
How we built it
The game is a simulation, written in TypeScript and drawn on an HTML canvas. It's seeded, so the same round plays out exactly the same way every time, which made testing a lot easier. Chrome handles the speech recognition, and RelAI talks back with the browser's voice.
For understanding what you say, common calls go through a quick keyword matcher, and a language model can step in for messier wording. Either way, every action goes through a validator before the bot touches it. Is that spot real, is it on our site, do we have smokes left? RelAI only talks to the bot through three simple calls (what do you see, here's a goal, did you finish it?), so in theory you could put it on top of any bot, including Riot's.
Being upfront: neither of us had built animations or a game simulation before, so we used Claude to write a lot of the TypeScript, the simulation and animation code, and the formatting and layout of the site. What we put our own time into was the AI agent itself, together. We figured out what RelAI should know about the round, what it's allowed to act on, when it should push back, and how its callouts stay tied to what it can actually see. We also designed the plays (Fast Plant, Fast Rush and Slow Push), the routes, the timing, and where the smokes and mollies go. Claude helped us move fast on the code, and the ideas and decisions behind RelAI are ours.
Challenges we ran into
Honestly, the biggest one was making talking matter. Our first version played out about the same whether you said anything or not, which kind of defeats the point. We had to change the game so defenders hold their angles and the bot only uses utility when you tell it to.
Speech recognition was its own adventure. Chrome kept hearing "B long" as "be long," which we found out while recording the demo. We also had to make sure the language model could never make up a spot you didn't say. And doing all of it in 24 hours with two people.
Accomplishments that we're proud of
We actually measured whether it works:
- Staying silent won 0 of 40 rounds. One smoke call won 29 of 40.
- We threw 15 trick inputs at it. RelAI let 0 bad actions through, while a basic "Copy." bot let 3 through.
- Every one of RelAI's callouts (247 of 247) matched an enemy it could really see.
Anyone can rerun these from the repo with npm run winrate and npm run evidence.
What we learned
A bot saying "copy" means nothing if it doesn't actually do it, so the checking matters more than the talking. We also learned that building our own sim was worth it, because it let us test every claim instead of just hoping. And we picked up a lot about how speech recognition breaks in real life.
What's next for RelAI
The big one is testing it with real players, to see whether practicing with RelAI makes your comms better in actual games. After that, handle more mishearings, add more sites and strategies, test the language-model side properly, and get it running on top of a smarter bot.
To be upfront about where it stands, it's our own 2D sim of one site with scripted enemies, it doesn't touch Valorant itself.
Talk to RelAI was created under Riot Games' "Legal Jibber Jabber" policy using assets owned by Riot Games. Riot Games does not endorse or sponsor this project.