Coral AI Based Game Engine

Real-time multiplayer games, built, populated and operated by AI.

Coral is one backend, one mobile shell and one AI pipeline sitting behind many live 1v1 games. AI authors the content, AI plays the opponent when no human is free, and AI reviews every change before it ships. Two games run on it from the same codebase — one live in the stores, one finished and awaiting release.

One substrate, many games

Games
word-clash · backgammon-clash · next title
AI layer
opponents · content · review · moderation
Mobile shell
20 shared screens · 16 theme packs · EN/TR
Framework
realtime · matchmaking · economy · social · admin
AWS
fargate · elasticache · cloudfront · s3 · cdk
2
games built on one engine
38
specialist AI agents in the pipeline
471
automated test suites
101
on-device UI test flows
227
live prompts across 16 categories
2
languages, end to end
400
concurrent players load-tested
~$35/mo
infrastructure at pilot scale

The problem

Real-time mobile games fail in three predictable places.

Coral exists because each of them is an AI problem wearing a game-development costume.

01

Empty lobbies

A 1v1 game needs two people in the same second. At launch, in a small market, or at 4am, there is only one — and a player who waits once rarely comes back.

02

The content treadmill

Session games burn through content faster than humans can write it, and every prompt has to work in every language you ship, with an answer set that is neither too tight nor too loose.

03

The boring 80%, rebuilt every time

Auth, matchmaking, economy, IAP, ads, social, moderation, admin tooling, store compliance. Every new title pays for it again before it can test whether the game is fun.

The AI layer

AI is not a feature bolted onto Coral. It is how the games run.

Four production systems, all live today in shipped apps — not a roadmap.

Opponents that pass for human

Every game carries a pool of AI personas across four skill tiers, each with its own accuracy jitter, personality, momentum arc and daily online schedule. Matchmaking hands you one within seconds when no human is free — so the lobby is never empty, in any market, at any hour.

  • Delay is composed from reading, thinking and typing time, not a constant
  • They answer the obvious thing first, because people do — then run out of ideas
  • Every deliberate mistake is re-checked against the same matcher that scores the game, so a miss cannot be silently auto-corrected into a hit
  • They come online on a schedule, accept your friend request minutes later, and never reveal themselves in a payload

An AI content pipeline with a human gate

The operator console runs the Claude Agent SDK to draft, translate and audit game content in place. Suggestions come back as a schema-validated patch a human approves — the model has no write access to the live bank, and every accepted change lands with an audit trail.

  • Multi-agent audits sweep the bank for semantic misfits and translation drift, with adversarial verification before anything is proposed
  • Content is versioned as a snapshot in git, so production and the repository can be diffed
  • Quality rubric and cut policy are written down and enforced, not vibes

Language understanding at the answer boundary

Scoring a free-text answer is the hard part of a word game. Coral resolves it server-side, per language, through a staged pipeline — exact, then a length-scaled fuzzy budget, then aliases, tokens and morphological variants — with a real-word guard that refuses to "correct" one valid word into a different one.

  • Turkish and English are separate lexicons; neither falls back to the other
  • Players see the word they typed, never a normalised key
  • The same matcher validates human answers and verifies AI mistakes — the two cannot drift apart

The engine builds itself

Coral is developed by an agent pipeline: specialist AI agents for planning, implementation, security review and release, driven by machine-readable rulebooks that encode every incident this platform has already survived. A small team ships at studio cadence because the review load is carried by machines.

  • 38 specialist agents and 23 workflow skills across the delivery lifecycle
  • Every pull request gets an AI code and security review before a human reads it
  • 471 test suites and 101 on-device UI flows gate the merge; infrastructure is code, deployed from CI with no long-lived credentials

The engine

Everything a live multiplayer title needs, already built and already in production.

A new game registers a module and inherits the rest. It is not a fork, and it is not a template — the two shipped titles run the same binary path.

Multi-game core

Game module registry, database per game, scoped cache keys, per-game store identity and economy tuning.

Real-time layer

WebSocket-first transport with polling fallback, unlimited reconnection, mid-match rejoin, cross-instance timer recovery.

Matchmaking

Language and skill matched, single-worker pairing with atomic claims, broadening criteria, AI fallback within seconds.

Economy

Triple currency, server-authoritative, atomic mutations with an immutable transaction ledger behind every balance change.

Monetisation

Apple and Google receipt verification, refund webhooks, replay-proof grants, and two ad networks switchable from config.

Social

Friends, challenges, invites with universal links, direct messages with moderation, presence, push notifications.

Mobile shell

20 shared screens, 16 theme packs, per-game branding and legal documents, one binary path across both games.

Operator console

Users, economy, content authoring and review, moderation queue, broadcasts, live metrics, audit log.

Remote control

Force update, maintenance mode, feature kill switches and ad provider changes without shipping an app build.

Localisation

English and Turkish end to end, including server-emitted strings, push copy and store surfaces.

Running on Coral

Two titles. Same backend, same shell, different game.

Architecture

Built on AWS, from the edge to the task definition.

Every environment is CDK infrastructure-as-code, deployed from GitHub Actions over OIDC with no long-lived credentials in the repository.

Players
iOS appAndroid appOperator console
Edge — TLS, static delivery
CloudFrontACMRoute 53S3 (OAC)
Application — realtime, authoritative
ALBECS Fargate · ARM64Node · Socket.IOautoscale 1–4
State
ElastiCache RedisMongoDB Atlascross-instance adapter
Operations
SSM Parameter StoreCloudWatch alarmsSNSAWS CDK

Measured, not estimated

Numbers from a production load test against the pilot stack — one 0.25 vCPU ARM64 task.

Concurrent players400
Peak CPU32.7%
Peak memory274 MB
Event-loop lag0.1 ms
Pilot cost~$35 / month

Designed for more than one task

Pairing is arbitrated by atomic operations in Redis, room emits cross instances through a streams adapter, and expired turn deadlines are swept by whichever task is alive. Scaling out is a number, not a rewrite.

Where AWS credits go

From a working pilot to a platform.

Coral runs a live multiplayer platform, with a second title built on it, for roughly the cost of a takeaway meal per month. Credits do not keep the lights on here — they buy the next order of magnitude.

  • 01

    Production topology

    The five-stack production environment is already authored in CDK — network, data, secrets, compute, CI. Credits take it from authored to multi-AZ and running.

  • 02

    Load headroom

    Test to ten thousand concurrent players instead of four hundred, and size the fleet on evidence rather than a curve fit.

  • 03

    Inference in our own account

    Evaluate Amazon Bedrock for the content and opponent pipelines so model inference runs beside the game data instead of leaving it.

  • 04

    More titles on the same substrate

    Each new game is a module, not another stack. The marginal cost of title three is the thing this architecture exists to prove.

Get in touch

Talk to us about Coral.

Partnerships, publishing, infrastructure programmes, or a look under the hood — we answer every message.