Sales technology — 2026
An agent platform that holds real customer conversations
Co-founded and led AI for a multi-agent platform that runs autonomous sales and customer engagement for automotive and real-estate clients.
Stack
The problem
Sales teams lose most of their pipeline to silence. Leads arrive, nobody answers within the hour, and the intent evaporates. The obvious fix — a chatbot — fails for a specific reason: a bot that cannot look up a real vehicle, book a real service slot, or hand over cleanly to a human is worse than no bot at all.
What we built
A multi-agent platform where each agent owns a narrow competence — qualification, inventory lookup, scheduling, escalation — and an orchestrator decides who speaks. Roughly fifty Rust services communicate over gRPC on an event bus, with the conversational state held outside the model so a conversation can survive a restart, a model swap, or a human taking over mid-thread.
The parts that mattered most were not the models:
- A tenant-scoped knowledge system so each client’s agents answer from that client’s own documents, inventory and pricing rather than from general training.
- An evaluation harness run against recorded conversations on every change, because “it seemed better in the demo” is not a release criterion.
- Deterministic handover. Every agent has a defined condition under which it stops and fetches a person.
Where it landed
The platform runs live customer engagement for automotive groups and a real-estate developer, integrated with dealer management systems for service booking. It also turned out to be the honest teacher about agentic systems: the failure modes are almost never the model’s reasoning, they are state, permissions, and the seams between services.