AI Flips 'Buyer Finds Supplier' to 'Supplier Finds Buyer' – Recruitment Agent Design

Background

Aliexpress lead generation suffers from high investment but low conversion, caused by three pain points: (1) valuable leads are not identified, (2) outreach efficiency is low, and (3) scaling is difficult.

Solution Overview

The system restructures the recruitment pipeline into two stages: intelligent lead screening & scoring, and AI‑driven voice interaction for first‑round outreach.

2.1 Intelligent Lead Insight (NL2SQL Agent)

An AI scoring model assigns a 0‑100 potential score and tags to each lead. Users describe selection criteria in natural language; the Agent translates the description into an SQL query that runs directly on ODPS, extracting the target leads without manual SQL writing or analyst involvement.

2.2 Intelligent Voice Interaction Engine

The interaction layer follows a three‑layer ASR + LLM + TTS architecture built on Alibaba Cloud voice services, Serverless functions, and the Qwen large model, orchestrated with LangChain.

Technical architecture : integrates Serverless compute, ASR, TTS, interruption detection, and VAD.

Scenario customization : supports custom greetings, scripts, and closing remarks.

Actor‑Tutor Dual‑Agent Logic

The Actor (fast) generates real‑time speech, handles interruptions, and detects user emotion. The Tutor (slow) uses stronger models (Qwen‑Max) and dynamic prompts to analyze full context, then feeds refined results back to the Actor, improving answer accuracy while preserving responsiveness.

Low‑Latency Optimizations (TTFT < 800 ms)

Two deployment options were compared: a centralized cross‑region setup (RT > 300 ms) and a same‑region Serverless deployment. The same‑region Serverless approach reduced network round‑trip time to < 100 ms. Additional JVM parameter tuning and Http thread‑pool reuse in LangChain4j supported high concurrency.

Multi‑path concurrency routes the fastest responding model (Turbo, Max, Plus) to the user.

Voice Tone Upgrade

Experiments with CozyVoice, Qwen3‑TTS, and voice‑cloned speech showed that a slightly imperfect cloned voice increased merchant engagement compared with a perfect synthetic tone.

AB Test Results

Live A/B testing reported roughly a 10 % increase in intent rate and higher interaction depth (≥5 rounds). 90 % of responses were under 1000 ms and all responses were under 1800 ms.

2.3 Intelligent Post‑Call Review

After each call, an intent‑recognition Agent scores merchant willingness based on dialogue, filtering high‑intent leads for follow‑up. A quality‑inspection Agent automatically analyses recordings and generates optimization suggestions, eliminating manual sampling.

Generalization of Voice Interaction

The ASR + Agent + TTS stack is modularized for other scenarios such as platform services, large‑scale promotions, supply expansion, and compliance notifications, enabling rapid configuration of voice agents for diverse business needs.