frontline
A 24/7 support agent and a marketing rep, streaming in real time, with a command center where a human can read every conversation the machines are having.
Most companies bolt a chatbot onto a website and call it AI support. It answers at machine speed and fails at machine scale, with nobody watching. The requirement is two-sided: customers get an agent that responds instantly and stays on-message; the business gets full visibility into what that agent says on its behalf.
A 24/7 support agent handles product questions and troubleshooting; a marketing rep handles the top of the funnel. Separate personas and objectives on one platform — a support answer and a sales answer should not share a prompt.
Replies stream token-by-token over server-sent events — no spinner, no wall-of-text dump.
Live stats and a full conversation browser: every exchange either agent has had, readable by the team it represents.
The API schema is the single source of truth: types and client hooks are generated from it, so the dashboard, the chat surface, and the backend cannot silently disagree about the shape of a conversation.
Stack: TypeScript monorepo — Express API over Postgres with a typed ORM, OpenAPI-generated schemas and query hooks, React + Vite front end, Claude behind both agents.
The hard part of customer-facing AI is not generating replies; it is letting the business supervise them. Every conversation either agent has ever had is stored and browsable, and the command center was built first-class, not as an admin afterthought. If a machine speaks in your name at 3am, you can read exactly what it said at 9.
A support answer optimizes for resolution; a sales answer optimizes for the next step. Blending them in one persona produces the worst of both — support replies that upsell, sales replies that troubleshoot. frontline keeps 2 agents with separate personas and objectives on one platform, so each conversation is judged by its own standard.
Our public demo runs the real interface against recorded conversation data, with streaming reproduced token-by-token — a faithful replay, clearly labeled.
Both agents ran against a live SaaS product surface: real conversations, live stats in the command center, streaming chat end to end. Built and operated by us.
Agents trained on your product and your voice, deployed on your site and channels, with the command center in your team's hands — and escalation to a human wired wherever you draw the line.