Inspiration
VALORANT's new replay system lets you rewatch every death from any angle. I got tired of doing exactly that: watching the same lost duel over and over with no way to change it. Would holding the angle have won it? Would the flash have worked? A replay can't tell you, and a generic bot isn't the player who beat you.
What it does
A 1v1 tactical shooter on VALORANT's rules (guns, shields, economy, spike, ranks; used with the hackathon's permission). Every round is recorded tick by tick on a deterministic server, so any moment can be rebuilt exactly.
- Replay: round list, kill timeline, per-round stats, four cameras, a 2D tactical map.
- What they saw: the rival appears only when the player you follow could actually see them.
- Coach: at any tick, the server plays 100 alternative futures (20 plans × 5 runs) in about a second and shows the best one in the replay.
- Take control: play from that exact tick. Your opponent repeats their real moves until you change the fight, then becomes a recreation of them.
- Every attempt is saved to the timeline: "YOU CHANGED IT · Originally lost · won 1 of 3".
The bot
Its role: your real opponent, recreated at the moment you lost to them.
- Out of fights: their own habits (where they go, how they move, where they look), copied from their recordings with k-nearest neighbours.
- In fights: an AI that aims and shoots at their rank, with their reaction time.
- New situations: when nothing in their recordings is close enough, a circuit breaker hands control to the AI.
- Traced: a live panel shows every decision:
GHOST→HANDOFF→MIX, and each copied move names the recorded moment it came from. - Fair: every bot sees only the player's fog-of-war view. No LLM: decisions run 32 times a second.
Evidence
Each policy played 6 matches against an AI at the person's rank. "Moves like them" is measured on a match it never trained on.
- Nipun
- Moves like them (clone / hybrid): 91% / 91%
- Rounds won (clone / hybrid / AI alone): 23% / 38% / 46%
- Pratul
- Moves like them (clone / hybrid): 23% / 21%
- Rounds won (clone / hybrid / AI alone): 9% / 25% / 44%
The hybrid keeps the person's movement and fights far better than a plain clone. That clone fired on only 22% of the ticks it could see you, which is why we built the hybrid. Replays and saved attempts rebuild to the exact recorded state.
How we built it
- Engine: deterministic TypeScript. A match is a seed plus inputs and replays to the same state hash.
- Server: authoritative Node at 32 ticks a second; the coach runs in a worker thread.
- Client: Unity 6 3D, also in the browser through WebGL.
- Process: built with AI coding agents (Claude Code, Kiro CLI), checked by tests, evals and play-testing. Every design decision is logged in DECISIONS.md.
Limits and next steps
- It only knows what a player has shown. With one short recording, the hybrid lost every round.
- Coach plans are scripted, not learned.
- Next: measure transfer (does rewinding a loss help you win that fight next time?), team modes, and forking from VALORANT's own replays.
Built With
- c#)
- claude-code
- cloudflare
- k-nearest-neighbours
- kiro-cli
- node.js
- three.js
- typescript
- unity-6-(urp
- vite
- vitest
- webgl
- websockets