PharmaLinea Voice Agent

Overview

PharmaLinea Voice Agent is a voice AI sales assistant for PharmaLinea, a B2B private-label food supplement manufacturer. The agent runs as a persistent side-panel widget on pharmalinea.com, holds natural voice conversations about the company's 36+ supplement products, walks prospects through clinical evidence, and qualifies inbound leads by collecting their contact details directly through the conversation.

Key Features

• Real-time voice conversation with low-latency turn detection and noise cancellation.
• Product knowledge across 36+ supplements in 11 categories, with clinical study data for substantiated products.
• In-conversation navigation: the agent opens relevant product and clinical pages on pharmalinea.com in the current tab; the widget lives in a persistent side panel, so the session continues across page changes.
• Voice-driven lead capture: when a prospect signals partnership interest, the agent gathers their contact details conversationally and submits them to the backend — no form, no interruption.

Technical Implementation

• Voice infrastructure: LiveKit Agents SDK for real-time audio between the user and the agent.
• Agent runtime (Python): voice pipeline composed of Silero VAD, Deepgram STT, an LLM, and ElevenLabs TTS, with LiveKit's multilingual turn detector and BVC noise cancellation.
• LLM: Google Gemini by default, with an Anthropic Claude alternative behind a single env flag.
• Widget (React 19 + Tailwind + shadcn/ui): side-panel agent session UI with aura audio visualizer and live transcript, persistent across page navigations.
• Backend (Express + TypeScript): issues LiveKit access tokens and exposes a lead-capture endpoint.
• Knowledge base: structured JSON for products, clinical studies, services, and FAQ — all loaded into the system prompt, no vector DB needed at this scale.
• Tool calls over RPC: the agent invokes browser-side navigation by sending RPC to the widget participant; server-side tools (get_product_details, get_clinical_study, submit_lead) run in the agent process.

Challenges & Solutions

The hardest part is keeping the conversation feeling natural while the agent is doing real work — looking up product details, navigating the user's browser, and deciding when to start collecting contact details. Solved with a clear split between server-side knowledge tools (fast, run inside the agent) and client-side navigation RPCs that don't tear down the live session, plus a system prompt that frames lead qualification as the agent's primary mission and instructs it to gather the prospect's details through the conversation itself.