Inspiration
I've played VALORANT, and one thing I've noticed is that getting better at the game isn't as simple as just playing more matches. Sometimes you lose a fight and don't really understand what you did wrong. Was your crosshair too low? Did you peek incorrectly? Were you standing in a bad position?
There are plenty of aim trainers, but most of them focus on shooting targets rather than teaching you how to make better decisions during an actual fight.
That's what inspired CLUTCHUP. I wanted to create something that doesn't just tell you what you're doing wrong, but actually helps you fix it.
What it does
CLUTCHUP is an AI-powered tactical FPS coaching simulator where players train against a personalized opponent in 1v1 matches.
Players start with a baseline assessment that measures things like crosshair placement, movement, shooting, peeking, and exposure to multiple angles. Based on the results, the system identifies weaknesses and recommends specific training exercises.
What makes CLUTCHUP different is that the bot doesn't just stand still or follow the same exact strategy every time. It can recognize repeated habits, like approaching through the same route or peeking the same angle, and adjust its behavior accordingly.
The game also includes guided demonstrations where the bot shows techniques before the player practices them.
After training, players can complete another assessment to compare their results and see where they improved.
CLUTCHUP also supports uploading gameplay recordings and reviewing YouTube videos. The video-review system includes player identity and spectator tracking so it can distinguish the user's gameplay from teammates being spectated. Automated visual feedback for uploaded videos is optional and requires an API connection.
How we built it
We built CLUTCHUP using JavaScript, HTML, CSS, Node.js, and browser-based 3D technology powered by Three.js.
We adapted existing open-source tactical FPS and aim-training components to create the playable training environment.
The coaching system collects gameplay measurements, such as shooting behavior, movement, positioning, and repeated approaches. These observations are saved locally and used to select training exercises.
For the adaptive AI, we implemented a rule-based tactical system that uses recorded player habits to change the opponent's behavior. For example, if a player repeatedly approaches through one route, the bot can choose a different position or angle.
We also built guided training demonstrations, baseline and reassessment systems, progress reports, gameplay video review, and player identity tracking.
OpenAI Codex was used to assist with development, implementation, debugging, and testing.
Challenges we ran into
One of the biggest challenges was making the AI opponent behave like an actual tactical player instead of a simple moving target. We had to consider navigation, shooting, cover, visibility, and how the bot should react to the player's actions.
Another challenge was making the coaching useful. We didn't want the system to give random advice. We wanted the feedback to come from things the game actually recorded.
Analyzing external gameplay recordings was also complicated. When a player dies in VALORANT, the camera switches to a teammate. Without proper detection, an AI could mistake that teammate's gameplay for the original player's. We worked on a system that identifies gameplay states and allows uncertain sections to be reviewed.
We also faced performance challenges because the entire game runs inside a browser.
Accomplishments that we're proud of
We created a playable 1v1 tactical training experience with four agent classes, 16 abilities, and an adaptive opponent.
We also built a coaching loop that takes a player from an initial assessment to a personalized training exercise and then a reassessment.
One feature we're especially proud of is the bot's ability to change tactics based on recorded player habits rather than treating every player the same way.
The latest build also passed 134 automated tests, according to our development testing. Live video-analysis accuracy still needs further testing.
What we learned
This project taught us a lot about game development, AI behavior, and how difficult it is to turn gameplay data into meaningful feedback.
We also learned that an AI bot doesn't necessarily have to be extremely difficult to be useful. Sometimes the best opponent is one that understands what the player needs to practice and creates situations that help them improve.
What's next for CLUTCHUP
In the future, we want to improve our gameplay analysis, expand spectator detection, introduce more advanced AI techniques, and make the opponent even more personalized.
We also want to explore training AI models using gameplay demonstrations so bots can eventually learn more complex tactical strategies.
Our long-term goal is to make CLUTCHUP a personal AI training partner that helps players improve their skills through actual gameplay rather than repetitive drills.
Built With
- artificial-intelligence
- computer-vision
- css3
- game-development
- gameplay-analysis
- git
- github
- html5
- javascript
- machine-learning
- node.js
- openai-api
- openai-codex
- three.js
- webgl
