Kaleido Platform architecture
Kaleido is a platform for casino games with two products that share one code base and one deployment:
- Kaleido Taxonomy tags games against a hierarchical taxonomy (theme, mechanics, features and so on). Tags come from people, from an in-app AI tagging agent, or from an external AI client over MCP. A person reviews every AI suggestion before it counts as confirmed.
- Kaleido GRE (the game recommendation engine) turns customer play into per-game affinities and blends them with tag similarity, co-play, popularity and newness into a ranked, explained list of games per customer.
This site describes how the platform is built: the services, what each one owns, how they talk to each other, and how each capability works step by step. Every page is written from the source tree; file paths point at the code that implements the behaviour.
How to read this site
| Section | What it answers |
|---|---|
| System context | Who uses the platform and which outside systems it touches |
| Containers | Every deployable unit, its port and its dependencies |
| Data ownership | Which service owns which Postgres schema or Redis key |
| Security | JWT for people, a shared service key between services, the gateway in between |
| Request flows | End-to-end sequence diagrams of the main journeys |
| Services | One page per service: responsibilities, API, internals, sequences |
| Capabilities | One page per capability: the algorithm or workflow behind it |
| Operations | Deployment, configuration, observability, failure drills, known limits |
The platform in one picture
Two recommendation paths
Kaleido GRE has two ways to get from play to recommendations. They share the scoring code and differ in where the customer's affinities come from.
| Path | Data flow | Freshness | Used for |
|---|---|---|---|
| Batch | Simulator or platform posts events to Activity (Postgres). Profile rebuilds affinities on demand. Recommender reads them over HTTP. | As fresh as the last profile build | Tuning, what-if, holdout accuracy tests |
| Live | Events go to Kafka, keyed by customer. Ingest keeps decayed per-game totals in Redis. Recommender reads Redis directly (?source=redis). | Seconds | Real traffic at thousands of events per second |
See Live streaming for the live path and Affinity for how both compute "how much does this customer like this game".
Technology
| Layer | Choice |
|---|---|
| Services | ASP.NET Core minimal APIs on .NET 10, EF Core with Npgsql |
| Database | PostgreSQL 17 with pgvector (one schema per service) |
| Streaming | Apache Kafka (KRaft) with Confluent.Kafka; JSON, Avro or Protobuf payloads |
| Live state | Redis 7.4 with a Lua script for atomic, idempotent updates |
| Front end | React 19, TypeScript, Vite, SignalR client |
| AI | Anthropic, OpenAI, OpenAI-compatible endpoints and Cursor agents, configured at runtime |
| AI interop | Model Context Protocol server (streamable HTTP) inside the Taxonomy API |
| Telemetry | OpenTelemetry (traces, metrics, logs) to a collector, then SigNoz |
| Packaging | Podman containers and podman-compose via deploy/deploy.sh |