Fair grading at game speed. An AI host that never plays favourites.

The challenge

Running a live pub-quiz game at a venue means making dozens of grading calls per session — and every disputed answer erodes player trust. Is The Fab Four an acceptable answer for The Beatles? Is a surname alone enough for a composer question? Human hosts are fast but inconsistent: different tables get different standards, and challenges eat into game flow.

Trivia & Tunes also needed to scale beyond what a single host could deliver. The platform serves both live venue games and home play sessions, in English and Norwegian, across multiple device types — and needed a single AI layer that could grade answers fairly across all of them without per-session human intervention.

On top of grading, the platform needed personality. Players expect a host with character — banter, catchphrases, commentary on wrong answers. That energy is hard to bottle in software, and impossible to sustain across dozens of simultaneous sessions without AI.

The solution

Gilligan Tech Inc. designed and deployed a multi-layer AI game engine on the Gilligan Tech Inc. AI Platform, covering three distinct capabilities: intelligent answer grading, AI host personality generation, and voice synthesis for spoken feedback.

The grading engine uses a 6-level strictness scale that venue hosts configure per game — from very lenient (accepts nicknames, partial phrases, close typos) to laser strict (exact match required). Before any LLM call, a local pre-grader handles exact matches, common synonyms, and typo variants to keep latency low and API costs controlled. Only ambiguous answers escalate to the LLM.

The host personality system — built around configurable AI personas like Finn — generates in-game commentary, round intros, and reaction lines in the host's voice. ElevenLabs voices premium audio assets; Piper TTS handles local fallback. Every generated line is cached so repeat moments are instant.

Implementation

1. LLM answer grading engine: A structured grading pipeline evaluates player answers against the correct answer and host-defined grading notes. The engine handles typos, nicknames, abbreviations, partial phrases, and surname-only answers. A 6-level strictness configuration lets hosts tune tolerance per game type — pub quiz vs. home game vs. competitive round.
2. Multi-provider LLM routing via LiteLLM: Grading requests route across Claude, GPT-4o, and AWS Bedrock Nova models with automatic primary/fallback logic. If the primary model is slow or unavailable, the fallback takes over within the same request cycle. All calls are logged to the ai_usage_events table with model, tokens, latency, and cost.
3. AI host personality system (HostBanter): Configurable host personas carry a name, vibe, catchphrase, language preference, and strictness override. The LLM generates contextual commentary — round intros, correct-answer reactions, wrong-answer lines — in character. Responses are cached per session to eliminate repeated LLM calls for stable moments.
4. Voice synthesis pipeline: ElevenLabs (eleven_multilingual_v2 for premium assets, eleven_flash_v2_5 for live in-game lines) renders host commentary as audio. Piper TTS provides a local fallback. Audio files are cached with TTL management so the same line is never synthesised twice.
5. AI bulk question generator: An admin tool generates trivia questions in bulk via GPT-4o, using a structured JSON schema to ensure consistent format. Integrates with Spotify, Apple Music, and YouTube for music metadata. Output is validated against the schema before entering the question bank.
6. Cost telemetry and admin dashboard: Every AI call — grading, commentary, voice, generation — is tracked with task type, model, tokens, latency, and estimated cost. The admin dashboard shows AI spend by model, task, and account, giving full visibility into per-venue AI economics. A credit-based billing layer ties AI spend directly to game sessions.

Results

Technology