ui + cs:go
your ai e-pal

opal

the all-in-one gaming experience with ai.

inspiration

people love gaming. but solo queue gets dull quick.

e-pal raised $30M from a16z because people will pay $30/hr just to have someone play games with them. 630k monthly visits. 18-24.

the demand is real. but the pattern doesn't change.

here's the thing nobody talks about: solo players login 3x less than people with regular teammates. riot knows this. valve knows this. epic games knows this.

they've tried everything:

better matchmaking (doesn't fix loneliness)
LFG systems (nobody uses them)
forced voice chat (instant toxicity)

none of it works because the problem isn't matchmaking. the problem is solo players don't have teammates.

so we built one.

what opal does

opal is your online friend who games with you.

no backseating. no coaching. not an overlay.

opal is in the lobby. holding angles. making rotations. calling out positions. acting like a real teammate.

the flow:

add opal on discord
ping opal in any channel
opal will join your vc
queue up and play!

that's it.

same brain, different games. we built a harness system - opal's intelligence is game-agnostic. add a new harness, opal plays a new game. cs:go today, tft tomorrow, valorant next week.

we're not building a CV bot. we're building AI teammate infrastructure.

how we built it

voice (livekit)

livekit handles the real-time comms. webrtc, VAD, the whole stack.

talk to opal like a teammate. "i'm flashing." "push A." "enemy mid."

opal responds back with customizable voice models.

~400ms round trip. fast enough for real callouts.

vision (overshoot + custom yolo)

different games need different approaches.

custom yolo

fps games are brutal on latency. every millisecond matters. we trained a custom yolov8 model specifically for enemy detection - lightweight, fast, does one thing well.

the model detects:

enemy player models
teammate positions
weapons on ground
bomb/objective status

runs at 60+ fps. fast enough for real combat.

tft, scribbl, slower games - overshoot

for games where you have time to think, overshoot handles the heavy lifting. we use it for rich context generation - understanding board states, parsing game UI, spatial reasoning for drawing.

overshoot feeds opal the full picture: "you have 3 star kai'sa, 47 gold, stage 4-2, lobby is weak." opal makes the call.

the ensemble approach: yolo for twitch reactions, overshoot for strategic context. best of both worlds.

the harness architecture

each game gets a harness. perception + action layer specific to that game.

the harness handles micro execution. pathfinding. steering. aiming. runs at 30fps independent of opal calls.

opal isn't in the hot loop. that's the key insight. the harness runs fast, opal provides strategy every 1-2 seconds.

challenges

voice latency. we started at 800ms, useless for callouts. livekit got us to sub 400ms.

vision speed. tried using heavyweight vision models for cs:go. 5fps. unplayable. ended up training a custom yolo - stripped down, enemy detection only. 60+ fps. overshoot handles the slower games where latency isn't life or death.

opal can't aim. brilliant at strategy, can't track a moving target at 144fps. solved by separating macro and micro - opal says "engage enemy A long", harness handles the actual aiming.

anti-cheat. geforce now sidesteps this. we're controlling a cloud VM, not touching the game process. game sees a normal windows machine.

what we learned

different games need different vision approaches. fps needs raw speed (custom yolo). strategy games need understanding (overshoot). the ensemble works.

opus 4.5 is genuinely good at game sense. didn't need to prompt engineer it into being a good teammate. the hard part was giving it eyes and hands.

everyone's building AI tools. nobody's building AI companions. but the demand is massive.