AI Agent Solution
AI agents wired into your systems, with handoff and audit trails
01Overview
A model that can chat is not an agent that can do a job. A production agent needs to know what data it may read, which APIs it may call and when to hand over to a human, and it must leave logs that explain every step.
We build agents on models available in mainland China, such as DeepSeek, Doubao and Qwen, choosing per task and keeping providers swappable. Typical forms are customer service agents, sales assistants, internal helpers and workflow automation, connected to CRM, ticketing, order or OA systems through tool calling, with human handoff, permissions and audit logs.
We start with one well-bounded, measurable use case, test it against real conversations, and widen scope and permissions step by step.
02The problems
Uncontrolled answers
A raw model on a support channel can invent policies or promise discounts that do not exist.
Disconnected from systems
If the agent cannot look up orders or update tickets, users still end up with a human.
Blurry permissions
Without clear data and action boundaries, an internal assistant can expose what it should not.
No way to measure
Without test sets and conversation reviews, quality is judged by gut feel.
03Architecture
Website chat, WeChat official accounts and mini-programs, WeCom, DingTalk or Feishu bots, all routed to one service.
Intent detection, dialogue state, tool planning and handoff rules.
RAG over your knowledge base plus least-privilege tools for orders, tickets and CRM.
DeepSeek, Doubao, Qwen and others behind one gateway for comparison, switching and metering.
Roles, output checks, full conversation and tool-call logs, human review and evaluation sets.
04Capabilities
Customer service agent
Answers from product and policy content, checks orders and shipping, and hands off with context.
Sales assistant
Qualifies visitors, suggests options and logs leads to CRM, leaving pricing to your sales team.
Internal assistant
Policy, process and project lookups in WeCom or DingTalk, scoped by department.
Workflow automation
Extracts data from email, forms or chat to create tickets and fill fields, with human sign-off on key steps.
Tool calling
Business APIs wrapped as tools with explicit parameters and permissions.
Human handoff
Triggered by keywords, sentiment, confidence or request, with full context for the agent taking over.
Audit logs
Inputs, retrieved sources, tool calls and outputs, replayable per session.
Evaluation
Regression test sets and reviews of escalated or poorly rated conversations.
05Typical scenarios
Typical situations — not specific client cases.
E-commerce support
For example, a store whose questions centre on shipping, sizing and returns, with disputes escalated.
B2B lead capture
For example, an industrial or software vendor capturing after-hours enquiries for next-day follow-up.
HR and IT helpdesk
For example, answering repeat questions on expenses and leave in WeCom and opening IT tickets directly.
Document extraction
For example, pulling fields from enquiry emails or photographed forms for human verification.
06Technology
- DeepSeek
- 豆包大模型（火山方舟）
- 通义千问（阿里云百炼）
- Python FastAPI
- PostgreSQL + pgvector
- 企业微信 API
- Function Calling
07FAQ
Why mainland Chinese models only?
For users in mainland China, data residency, stability and compliance matter; DeepSeek, Doubao and Qwen cover most support and office tasks. Providers stay swappable.
How do you stop it making things up?
Grounded answers with sources, forced handoff on pricing and commitments, output checks, and regression testing. Handoff and audit logs remain essential.
Can the agent write to our systems?
We start read-only, then allow low-risk writes with logging and require human approval for high-risk ones.
Related development services
Industries
Tell us about your project
Leave your details and we’ll reply within one business day. We start with your needs, then a proposal; pricing is discussed directly.
- Reply within one business day
- Free initial conversation, no hard sell
- Handover with admin and docs