Prototype
Gallery Side
Exploring AI-driven cricket commentary with four distinct character voices, a stats firewall, and a silence-first speak gate.
Visit Gallery SideProblem
AI-generated sports commentary either invents statistics it was never given, speaks on every ball regardless of whether the moment earned it, or loses character identity after a few exchanges. The result sounds like a single generic voice narrating, not a panel disagreeing about what they just watched.
Exploration
A linear ball-by-ball Python pipeline where each module has one contract. A BallEvent normalises deliveries across data sources. A stats firewall computes derived numbers and hard-limits what the LLM can cite — it never fabricates a figure. A speak gate defaults to silence and fires only when a delivery crosses trigger thresholds. Four fixed Bangladeshi hosts (optimist, doomer, analyst, realist) rotate in pairs every five overs, with a friction rule that ensures the two active characters always carry some tension. Character identity lives in per-character config files so voice is tunable without touching code. The LLM client is injected, so the real Anthropic API and a scripted fake both satisfy the same protocol. 107 tests run fully offline.
Audience
Bangla cricket fans following Bangladesh matches. Secondarily, builders interested in character-consistent multi-agent dialogue systems and the silence-default design pattern.
Possible Model
No commercial model yet. The system is a research prototype exploring the design constraints for character-consistent AI commentary at the delivery level.
Risks
Hallucination risk is structurally bounded by the stats firewall but not eliminated — if the data source sends a wrong number, the firewall repeats it faithfully. Character consistency degrades over long matches as the context window rolls. The silence default makes the system feel considered but also means a cautious speak gate leaves long quiet stretches on low-drama matches. Bangla voice quality depends heavily on LLM Bangla capability, which is uneven.
Founder Notes
The silence default was the hardest design decision to commit to and the most important. A system that speaks on every ball is a monologue with character costumes. The speak gate earning its fire is not a filter on output — it is what makes the output mean something. The stats firewall invariant (LLM only rephrases numbers from the facts dict, never originates them) is the other non-negotiable. Without it, the commentary reads confident and wrong.