The problem
Complex enterprise deals increasingly fail because buying groups cannot align around a decision. CRM systems capture activity, but they do not necessarily help a seller understand the decision system itself: who matters, what each stakeholder values, where confidence is missing, and what must happen next.
What I built
Symphony became the environment in which I learned to turn that problem into working AI. It now spans contextual coaching, deal intelligence, stakeholder analysis, multi-source research, meeting preparation, realtime roleplay, voice interaction and institutional tender knowledge.
One example is an event-driven meeting agent that detects an upcoming meeting, identifies the people involved, assembles the relevant deal and stakeholder context, generates a meeting-specific briefing and distributes it to the appropriate internal participants.
Architecture in practice
Symphony is deliberately not one model wrapped in one interface. Different workloads use different paths depending on whether the job is extraction, synthesis, research, coaching, retrieval or live conversation.
Switchable roleplay voice architectures
The roleplay environment also lets me compare two different conversational architectures against the same personas and evaluation logic. One path uses an integrated OpenAI realtime model. The other separates the conversational model from speech and uses Deepgram Flux for the voice layer.
I have kept both behind a switch rather than forcing an early winner. That allows latency, turn-taking, conversational quality and cost to be tested without redesigning the underlying roleplay workflow.
Industrial AI translation
The domain is sales, but the underlying pattern is not. Symphony takes an event or question, assembles the relevant operational context, reasons over it and turns the result into something a person can act on.
What became reusable
As the system grew, some capabilities stopped being "features" and became reusable building blocks. The research engine was extracted into a shared package. Tender intelligence became its own evidence-and-confidence problem. Roleplay required state, persona control and separate evaluation. Assurance work emerged because increasingly capable agents needed stronger controls.