WorkflowTime

Our agent is a voice-first LLM Debate Engine: the user speaks a dilemma, a main “orchestrator” agent structures the problem, spins up multiple specialized debate agents, and guides them toward a compromise the user can influence in real time. The working prototype is stable enough for live demos, with a full STT → LLM → TTS → email pipeline, low-latency voice interaction, and automatic summarization wired into production-ready workflows. Technically, it coordinates multimodal input/output (voice + text), tool orchestration (Deepgram, ElevenLabs, n8n), and dynamic agent creation from a single user prompt. The result is an unusually transparent, interactive decision assistant instead of a one-shot answer bot. This pattern can support real-world use cases like career choices, product prioritization, and internal decision-making in teams, where trade-offs and debate matter more than a single “correct” answer.

Theme alignment & technologies

The project aligns with the “browsers, voices, clouds, and tools as agents” theme by turning each layer into part of the decision-making system:

Voices → agents: Deepgram handles real-time speech-to-text so spoken dilemmas become structured inputs; ElevenLabs gives each debate agent its own synthetic voice, so different perspectives literally “speak” to the user.

Browsers → interface agent: A React + Vite frontend acts as the user’s control surface, streaming debate state, handling user interventions, and keeping the experience responsive in the browser.

Clouds → orchestration layer: The LLM debate logic and the ElevenLabs TTS microservice run as independent cloud services, so voice generation can scale separately without blocking the core debate engine.

Tools → automation agent: An n8n workflow listens for final debate states and automatically generates and emails the structured summary, turning the reasoning process into a reusable, automatable tool.

Technologies, frameworks, and tools used

Frontend: React 18 (component-based UI, hooks like useState, useEffect), Vite dev server/bundler (fast HMR, smooth DX)

LLM layer: Hosted chat-completion LLM (LLM debate orchestrator + debate agents)

Speech stack: Deepgram API (speech-to-text), ElevenLabs API (text-to-speech, multiple voices), exposed via a dedicated microservice

Automation & integration: n8n (email delivery and workflow automation around debate summaries)

Architecture: Multi-agent debate orchestration (main agent + dynamically created debate agents), microservice separation for audio so it can scale independently of the core logic.