Rillius — player-intelligence & game-recommendation platform

How we built it

Delivered in deliberate phases.

1. 01

   Per-tenant models, proven isolation

   Every customer gets its own models, trained on its own data and scored in its own isolated world, with isolation pushed into the database via row-level security rather than trusted to application code alone. Trained bundles are cached against each tenant's data so scores stay fresh without being recomputed on every page view.

2. 02

   PII-free ingestion at the boundary

   Every feed is defined by a versioned contract and passed through a guard that rejects any batch carrying personal data — matching both field names and value patterns — so the platform never stores raw PII. A dry-run endpoint lets a company's engineers validate payloads without landing anything.

3. 03

   Calibrated, honest predictions

   Lifetime value ships as a calibrated range, not a bare number, using a hurdle model for the heavy-tailed, mostly-zero distribution and conformal prediction for distribution-free confidence bands. A tenant with too little data is told so, rather than scored through a fabricated number.

4. 04

   Real-time recommendations and decisioning

   An offline collaborative-filtering model learns taste and an online contextual bandit decides what to surface now, exploring enough to avoid an echo chamber. A decisioning layer ranks the next-best action per player by a transparent expected value, with wellbeing checks treated as non-negotiable imperatives.

5. 05

   Governance and monitoring

   Model changes are governed rather than free-text edits, health is snapshotted per scoring run with drift flags, and a per-action uplift flywheel stays dormant — labelled an assumption — until real treated-and-control outcomes let it learn.