AI · Real-time LLM orchestration
LIVE DEMO
Liga Arena
A real-time dual AI sportscaster for a live Kick game: a fault-tolerant multi-model LLM cascade generates contextual play-by-play, with Spanish neural TTS and speculative pre-synthesis for ~0ms perceived latency.
- Multi-model LLM cascade (Groq Llama 3.3 70B → Cerebras Qwen/Llama → Gemini) with kind-aware routing per event — no single point of failure
- Production reliability decorators: Circuit Breaker, token-bucket Rate Limiter, Timeout, header Quota Tracking and CAG cache — composed via DI
- Speculative inference: pre-generates winner candidates (LLM + TTS) ~15s before climax → ~0ms perceived latency
- Semantic anti-repetition with local embeddings (transformers.js MiniLM) + pgvector cosine — no external similarity API
- Defensive prompt engineering: 7-bucket output validator (anti-copyright, anti-hallucination) + prompt-injection guards
Groq · Llama 3.3 70BCerebras · Qwen 235BGemini
XTTS v2 TTStransformers.jspgvector
Circuit BreakerSpeculativeTypeScript
Live demo on request