| Technical details for those interested: The Neuron framework handles the orchestration complexity - it's designed around biological neural principles for resilient reasoning and arbitration between agents. This lets the agents maintain distinct personalities while still having coherent debates. The hardest part was training "personalities" that would actually disagree. Early versions had all agents converging on the same picks. I ended up training each LoRA on different analyst archetypes (contrarian, stats-heavy, narrative-focused, etc.) with curated datasets for each. For voice streaming, ElevenLabs has great emotion but latency issues, so I mix in OpenAI TTS for snappier responses. Currently building a queueing system to pre-generate common debate segments. Cost breakdown: ~$0.02 for model inference, ~$0.02 for TTS per debate. At scale this could get expensive, but caching common player comparisons helps a lot.
The streaming angle has been interesting - we're going live on Twitch/YouTube during NFL games starting Sept 5. The idea is to have the agents debate in real-time as games unfold, which creates a natural funnel to the product. Happy to answer questions about the multi-agent orchestration, LoRA training process, or the business side of fantasy sports tools. |