以结果为起点Start from the outcome
启动前与您共同定义可度量的运营指标，例如处理时长、准确率、人工投入或单次成本，并记录基线。Before we build, we agree with you on measurable operating metrics — handling time, accuracy, manual effort or cost per task — and record the baseline.
以前线部署工程（FDE）模式交付Forward Deployed Engineering (FDE)
让大模型
成为生产力。Put large models to work.
前线部署工程师（FDE）与您的团队并肩工作：从企业知识库、智能体（AI Agent）到系统集成与私有化部署，交付真正投入生产的系统，并按约定的业务指标验收。 Forward deployed engineers work side by side with your team — from knowledge bases and AI agents to systems integration and self-hosted deployment — delivering systems that run in production, accepted against business metrics you agree on up front.
示例 · 供应链智能体Demo · Supply-chain agent
演示对话，所示数据均为虚构。Illustrative conversation; all data shown is fictional.
交付理念Our approach
从试点，
走向生产。From pilot
to production.
不少 AI 试点止步于演示：数据无法接入真实业务系统，效果缺少统一的衡量标准，上线后也无人持续运营。我们从项目启动之初就按生产标准构建，让每一步都有可度量的依据。Many AI pilots stall at the demo: the data never reaches real systems, no one agrees on what “good” looks like, and no one runs the system after launch. We build to production standards from day one, so every step rests on something you can measure.
生产级工程Production-grade engineering
评测、权限、系统集成、监控与回滚，从原型阶段起就按生产标准设计。Evals, permissions, systems integration, monitoring and rollback — designed to production standards from the first prototype.
能力移交Capability transfer
移交源代码、评测集与运行手册，培训您的团队，并协助建立 AI 卓越中心（CoE）。We hand over source code, eval sets and runbooks, train your team, and help you set up an AI Center of Excellence (CoE).
解决方案能力Capabilities
从规划到运营，
一支团队完成。From strategy to operations,
one team.
六类能力覆盖大模型应用从规划到运营的每个环节：可以整体交付，也可以按需组合，与您已有的平台和团队协同。Six practices cover every stage of an LLM application, from planning to operations. Engage them end to end, or combine them with the platforms and teams you already have.
AI 咨询与规划AI advisory & planning
盘点数据、系统与组织基础，筛选值得优先推进的场景，测算投资回报，并制定分阶段路线图。We assess your data, systems and organization, shortlist the use cases worth funding, model the return, and lay out a phased roadmap.
AI 成熟度评估AI readiness assessment场景路线图Use-case roadmapROI 测算ROI modeling模型与平台选型Model & platform selectionAI 治理框架AI governance framework
智能体与应用工程Agents & application engineering
构建能检索知识、调用工具并完成多步任务的智能体，通过 API 与 MCP 接入 ERP、CRM、OA 等业务系统。We build agents that retrieve knowledge, call tools and complete multi-step tasks, connected to your ERP, CRM and office systems through APIs and MCP.
企业知识库与 RAGKnowledge base & RAG智能体开发Agent development工作流编排OrchestrationMCP 连接器MCP connectors智能文档处理（IDP）Intelligent document processing (IDP)智能问数（ChatBI）Conversational BI
模型工程与数据Model engineering & data
以提示词与上下文工程、微调与后训练提升效果，以蒸馏与量化降低成本；评测集贯穿始终，衡量每一次改进。Prompt and context engineering, fine-tuning and post-training raise quality; distillation and quantization lower cost. Eval sets measure every change along the way.
提示词与上下文工程Prompt & context engineering微调与后训练（SFT / LoRA / DPO）Fine-tuning (SFT / LoRA / DPO)蒸馏与量化Distillation & quantization评测集建设Eval set construction数据治理Data governance
部署与推理Deployment & inference
在您的机房、专有云或混合云中部署模型与应用，并按首词元时延、吞吐与成本优化推理服务。Models and applications deployed in your data center, a dedicated cloud or a hybrid setup, with inference tuned for time to first token, throughput and cost.
私有化部署Self-hosted (on-prem) deployment专有云与混合云Dedicated & hybrid cloud大模型一体机LLM appliance推理优化（vLLM / SGLang）Inference tuning (vLLM / SGLang)AI 网关与多模型路由AI gateway & model routing信创适配Xinchuang-ready
安全、合规与治理Security, compliance & governance
从输入输出护栏到权限审计，安全要求在设计阶段就内置于系统；我们也会协助完成上线所需的合规工作。From input and output guardrails to access audits, security is designed into the system from the start — and we help with the compliance work a launch requires.
护栏与内容安全Guardrails & content safety红队测试Red teaming权限与审计Access control & audit数据脱敏Data masking备案与登记支持Filing & registration support生成内容标识AI content labeling
持续运营（LLMOps）Operations (LLMOps)
上线后持续评测、监控与优化：每次变更先过质量门禁，词元（Token）用量与成本清晰可控。After launch we keep evaluating, monitoring and improving. Every change passes a quality gate, and token usage and cost stay visible and under control.
持续评测与质量门禁Continuous evals & quality gates可观测性Observability词元成本治理Token cost governance坏例闭环Error analysis效果复盘Value reviews托管运维Managed operations
应用场景Use cases
按业务职能，
找到您的场景。Find your use case,
function by function.
以下按业务职能整理了常见的切入场景。任何一个场景，都可以从一次基于您真实数据的联合创新工作坊开始。Common starting points, grouped by function. Any of them can begin with a co-innovation workshop on your own data.
智能客服与坐席辅助AI customer service & agent assist
结合知识库与订单系统自动应答；复杂问题转人工时，自动生成会话摘要与建议回复。Answers routine requests from the knowledge base and order system, and hands complex cases to human agents with a summary and a suggested reply.
知识库问答Knowledge Q&A工单摘要Ticket summaries坐席辅助Agent assist
智能质检Conversation QA
对通话与在线会话全量质检，识别违规话术、情绪风险与服务改进点，替代人工抽检。Reviews every call and chat for non-compliant scripts, sentiment risk and coaching points — instead of manual sampling.
语音转写Transcription合规检测Compliance checks
营销内容生成Marketing content generation
按品牌规范批量生成多渠道文案与素材，内置审核流程与生成内容标识。Generates on-brand copy and assets for every channel at scale, with review and AI-content labeling built in.
品牌规范Brand rules多语种Multilingual内容审核Review flow
销售助手Sales copilot
自动整理客户画像与会议纪要，生成跟进计划和方案初稿，并回写 CRM。Builds account briefs and meeting notes, drafts follow-ups and proposals, and writes back to the CRM.
CRM 集成CRM integration会议纪要Meeting notes方案生成Proposal drafts
研发智能体Coding agent
私有化部署的代码补全、评审与测试生成，基于内部代码库检索，代码不出内网。Self-hosted code completion, review and test generation grounded in your repositories — code never leaves your network.
代码评审Code review测试生成Test generation私有化Self-hosted
IT 运维助手AIOps assistant
聚合日志、指标与告警，自动定位根因，生成处置建议与变更记录。Correlates logs, metrics and alerts to find root causes, then drafts remediation steps and change records.
告警分析Alert triage根因定位Root cause
合同审查Contract review
抽取关键条款，标注风险与偏离项，生成审查意见初稿，经法务复核后定稿。Extracts key clauses, flags risks and deviations, and drafts review comments for legal to finalize.
智能文档处理Intelligent document processing风险标注Risk flags人工复核Human review
费用审核Expense audit
识别发票与单据，对照报销制度自动审核，异常单据附理由转人工处理。Reads invoices and receipts, checks them against policy, and routes exceptions to a reviewer with reasons.
票据识别Receipt OCR制度校验Policy checks
财报解读Financial statement analysis
抽取财务报表关键数据，计算同比与环比变化，生成附原文出处的分析摘要。Extracts key figures from financial statements, computes year-over-year and sequential changes, and writes a summary that cites its sources.
报表抽取Statement extraction指标计算Ratio analysis引用溯源Cited answers
招聘助手Recruiting assistant
按岗位要求解析与筛选简历，生成面试问题与评估表；录用决定始终由招聘团队作出。Parses and screens résumés against the job requirements and drafts interview questions and scorecards — hiring decisions stay with your team.
简历解析Résumé parsing面试评估表Interview scorecards人在回路Human in the loop
员工自助服务Employee self-service
结合人事制度与 HR 系统数据，按员工身份回答假期、薪酬与福利问题，并可直接发起申请。Answers leave, payroll and benefits questions from HR policy and system data, scoped to each employee — and can file the request for them.
制度问答Policy Q&AHR 系统集成HR system integration权限感知Permission-aware
供应链智能体Supply-chain agent
调用 ERP、MES 等系统工具，自动完成库存核查、排产评估与交付确认。Calls ERP and MES tools to check inventory, assess schedules and confirm delivery dates.
工具调用Tool calling多系统集成Multi-system
设备运维助手Maintenance copilot
结合设备手册、维修记录与实时告警，给出排障步骤与备件建议。Combines manuals, repair history and live alarms into step-by-step fixes and spare-part suggestions.
故障诊断Diagnostics经验沉淀Know-how capture
智能问数Conversational BI
用自然语言查询经营数据，自动生成图表，并对指标异动做归因分析。Ask business questions in plain language and get charts plus root-cause analysis of metric changes.
ChatBI指标归因Root-cause analysis
经营周报自动生成Automated weekly business reviews
定时汇总多个系统的经营数据，生成图表、异动说明与下周关注事项，经负责人确认后发布。Pulls figures from multiple systems on a schedule and drafts charts, variance notes and next week's priorities — published once the owner signs off.
定时生成Scheduled runs指标异动Variance notes人工确认Owner sign-off
企业知识助手Enterprise knowledge assistant
连接文档、制度与业务系统，按权限回答问题，每个答案都可追溯到原文。Connects documents, policies and systems to answer within each user's permissions — every answer cites its source.
权限感知Permission-aware引用溯源Cited answers
深度研究与报告Deep research & reports
多源检索并交叉核验，生成结构化研究报告与图表，每个结论都标注来源。Searches across sources, cross-checks what it finds, and produces a structured report with charts — every finding linked to its source.
多源检索Multi-source search交叉核验Cross-checking报告与图表Reports & charts
会议纪要与行动项Meeting notes & action items
转写会议录音，生成结构化纪要，提取行动项与负责人，并同步到任务系统。Transcribes the recording, writes structured minutes, pulls out action items and owners, and syncs them to your task tracker.
语音转写Transcription行动项提取Action items任务同步Task sync
行业方案Industries
每个行业，
都有自己的关键场景。Every industry has
its own critical workflows.
我们按行业梳理典型场景、常见数据来源与合规要求，帮助您更快确定起点。We map typical use cases, common data sources and compliance requirements industry by industry, so you can choose a starting point sooner.
金融Financial services
银行 · 证券 · 保险Banking, securities and insurance
- 智能投研AI-assisted investment research
- 信贷报告生成Credit memo drafting
- 研报合规审查Research report compliance review
政务与公共服务Government & public sector
政务服务 · 城市治理 · 公共事业Public services, city management and utilities
- 12345 热线工单分拨12345 government hotline ticket routing
- 公文辅助写作Official document drafting
- 政策问答与导办Policy Q&A and service guidance
制造Manufacturing
离散制造 · 流程制造Discrete and process manufacturing
- 设备故障诊断与预测性维护Fault diagnosis & predictive maintenance
- 工艺知识库Process knowledge base
- 质量检验Quality inspection
能源与电力Energy & utilities
发电 · 电网 · 油气Power generation, grid, oil and gas
- 调度辅助Dispatch decision support
- 巡检报告生成Inspection report drafting
- 规程问答Operating-procedure Q&A
医疗与生命科学Healthcare & life sciences
医院 · 医药 · 医疗器械Hospitals, pharma and medical devices
- 病历文书生成Clinical documentation drafting
- 医保审核Medical insurance claim review
- 医学文献检索Medical literature search
零售与消费品Retail & consumer goods
零售 · 电商 · 消费品牌Retail, e-commerce and consumer brands
- 智能导购AI shopping assistant
- 商品内容生成Product content generation
- 会员运营Member engagement
汽车与出行Automotive & mobility
整车 · 零部件 · 出行服务Automakers, suppliers and mobility services
- 研发知识库R&D knowledge base
- 售后诊断After-sales diagnostics
- 座舱语音交互In-car voice assistant
电信运营商Telecommunications
移动 · 宽带 · 政企业务Mobile, broadband and enterprise services
- 网络运维Network operations
- 智能客服AI customer service
- 精准营销Targeted marketing
教育Education
高校 · 职业教育 · 教育科技Universities, vocational schools and edtech
- 智能助教AI teaching assistant
- 个性化学习Personalized learning
- 教研内容生成Course content generation
前线部署工程（FDE）Forward Deployed Engineering
嵌入式交付，
按结果验收。Built inside your team.
Accepted on results.
前线部署工程师（Forward Deployed Engineer，FDE）兼具软件工程能力与大模型应用经验。他们进入您的业务一线，观察真实流程，与领域专家共同构建，直到系统在生产环境中稳定运行，并对约定的业务指标负责。Forward deployed engineers (FDEs) are software engineers with hands-on experience building on large models. They work on your front line, observe the real workflow and build alongside your domain experts until the system runs reliably in production — accountable for the business metrics you agreed on.
01
携手共建Build together
FDE 进入您的业务一线，与领域专家组成联合交付团队，直接在您的系统和数据上构建。FDEs join your front line and form one delivery team with your domain experts, building directly on your systems and data.02
攻坚生产难题Take on hard problems
专注现成产品解决不了的问题：复杂的系统集成、权限与数据治理，以及效果、时延与成本之间的取舍。We focus on what off-the-shelf products can't solve: complex integrations, permissions and data governance, and the trade-offs between quality, latency and cost.03
沉淀可复用资产Make it reusable
把验证过的方案沉淀为智能体模板、连接器与评测集，后续场景在此基础上构建，交付周期随之缩短。Proven solutions become agent templates, connectors and eval sets, so each new use case starts from what already works — and ships sooner.五个阶段，
每一步都有通过标准。Five stages,
each with clear exit criteria.
进度可检验，投入可控制：每个阶段结束时，由书面的通过标准决定是否进入下一步。Progress you can verify and spend you can control: at the end of each stage, written exit criteria decide whether to move on.
-
场景评估Discover
1–2 周1–2 weeks访谈与流程走查，用价值—可行性矩阵排定场景优先级。Interviews and process walkthroughs; a value–feasibility matrix ranks the candidates.
- 确定首批 1–3 个场景First 1–3 use cases selected
- 基线与成功标准经双方确认Baseline and success criteria signed off
-
共创验证（PoC）Prove (PoC)
2–4 周2–4 weeks用您的真实数据构建原型，建立评测集与效果基线。Build a prototype on your real data and set up the eval set and quality baseline.
- 评测集指标达到约定阈值Eval metrics clear the agreed threshold
- 试点范围与系统接入方案确定Pilot scope and integration plan agreed
-
试点上线（MVP）Pilot (MVP)
6–12 周6–12 weeks在一个部门或区域接入真实系统与权限，灰度运行。Connect real systems and permissions in one team or region, and roll out gradually.
- 业务 KPI 相对基线达标Business KPIs hit target against the baseline
- 安全评审通过Security review passed
-
规模化推广Scale
按场景推进Use case by use case通过上线门禁后正式投产，再把验证过的方案复制到更多部门。Go live once the launch gates are cleared, then roll the proven solution out to more teams.
- 压测、红队测试与合规检查（等保、备案、标识）完成Load tests, red teaming and compliance checks (MLPS, filing, labeling) done
- SLA 与运行手册生效SLA and runbooks in effect
-
持续运营（LLMOps）Operate (LLMOps)
长期Ongoing回归评测、坏例闭环、词元成本治理与模型升级。Regression evals, error analysis, token cost governance and model upgrades.
- 每次变更先过质量门禁Every change passes a quality gate
- 按季度复盘业务价值Quarterly value reviews
周期为典型参考，实际周期取决于场景复杂度以及数据与系统接口的就绪程度。Durations are typical; actual timelines depend on scope and on how ready your data and system interfaces are.
交付物You keep
源代码Source code
评测集Eval sets
架构文档Architecture docs
运行手册Runbooks
监控看板Dashboards
培训材料Training materials
FDE 与传统项目外包FDE vs. traditional outsourcing
同样是交付，
衡量标准不同。Same project.
A different measure of success.
外包按需求说明书交付功能；FDE 对业务结果负责，并在过程中把能力留在您的团队。Outsourcing delivers features to a spec. FDE is accountable for business results — and leaves the capability with your team.
| 对比项Aspect | 传统项目外包Traditional outsourcing | 前线部署工程（FDE）Forward deployed engineering |
|---|---|---|
| 交付目标Goal | 按需求说明书交付功能Features delivered to a spec | 达成约定的业务指标Agreed business metrics achieved |
| 团队形态Team | 供应商团队按需求开发，与业务一线距离较远A vendor team builds to requirements, far from the front line | 工程师嵌入业务一线，与领域专家组成联合团队Engineers embedded on the front line, in one team with your domain experts |
| 应对不确定性Handling uncertainty | 范围变化需走变更流程，调整周期较长Scope changes go through change requests and take time | 短周期迭代，由评测集与基线数据决定方向Short iterations; eval sets and baseline data set the direction |
| 交付物Deliverables | 系统与项目文档A system and project documents | 生产系统，以及源代码、评测集、运行手册与监控看板A production system, plus source code, eval sets, runbooks and dashboards |
| 上线之后After launch | 验收即结项，运维通常另行采购Sign-off ends the project; operations are usually bought separately | 持续评测与运营，按季度复盘价值，并把能力移交给您的团队Continuous evals and operations, quarterly value reviews, and capability handed to your team |
适合 FDE 的问题Where FDE fits
FDE 尤其适合以下几类问题：FDE is built for problems like these:
- 关键环节效率不高，但不确定如何用 AI 改进A critical workflow is slow, and it's unclear how AI should help
- 场景超出现有模型或产品的可靠能力，需要工程攻坚The use case goes beyond what current models or products handle reliably
- 同一问题在多个团队反复出现，值得做成可复用的方案The same problem keeps surfacing across teams and deserves one reusable solution
合作前提What we need from you
以下条件就绪，项目推进会更顺利：Projects move faster when these are in place:
- 明确的业务目标与负责人A clear business goal and an accountable owner
- 领域专家的持续投入Ongoing time from your domain experts
- 相关系统与数据的访问权限Access to the systems and data involved
- 可行的度量方法A practical way to measure results
技术架构Architecture
模型无关，
按需部署。Model-agnostic.
Deployed on your terms.
采用模型无关（Model-agnostic）架构，适配已备案的国产大模型与主流开源模型，可在您的环境内私有化部署。模型更新换代时，应用与数据层无需重建。A model-agnostic architecture works with domestic models that have completed generative-AI filing and with leading open-source models, and can run self-hosted in your environment. When models change, your applications and data layer don't have to be rebuilt.
- AI 网关统一鉴权、限流、计量与故障切换One AI gateway for authentication, rate limits, metering and failover
- 检索遵循原系统权限，答案可追溯到原文Retrieval honors source-system permissions, and every answer cites its source
- 高风险操作保留人工确认（人在回路）High-risk actions keep a human in the loop
贯穿各层的 LLMOpsLLMOps across every layer
评测Evals
可观测性Observability
成本治理Cost governance
渠道与应用Channels & appsL5
企业微信WeCom钉钉DingTalk飞书Feishu门户Portals业务系统Business systems
智能体与编排Agents & orchestrationL4
智能体Agents工作流Workflows人在回路Human in the loopMCP 连接器MCP connectors
模型服务Model servingL3
AI 网关AI gateway多模型路由Model routing国产大模型Domestic models开源模型Open-source models微调模型Fine-tuned models
数据与知识Data & knowledgeL2
RAG向量检索Vector search数据治理Data governance权限同步Permission sync
算力底座ComputeL1
安全与合规Security & compliance
安全，
从架构开始。Security starts
with the architecture.
数据存放在哪里、谁可以访问、智能体能做什么——每一项都可配置、可审计。Where your data lives, who can see it and what each agent may do — every part is configurable and auditable.
数据主权Data sovereignty
支持私有化部署。数据不出域，指以下三点均在您的环境内完成：Deploy self-hosted and your data stays in your domain — all three of these happen inside your environment:
- 存储：您的机房或专有云Storage: your data center or dedicated cloud
- 推理：您环境内的 GPU，不调用公网模型 APIInference: GPUs inside your environment, no public model APIs
- 流转：不经过外部节点Data flows: no external nodes
访问控制Access control
对接企业统一身份认证，沿用您现有的组织架构与系统权限。Plugs into your identity provider and inherits the org structure and permissions you already manage.
- 单点登录（SSO）Single sign-on (SSO)
- 基于角色的访问控制（RBAC）Role-based access control (RBAC)
- 检索结果遵循原系统权限Retrieval honors source permissions
- 访问与操作审计日志Access and action audit logs
智能体权限Agent permissions
为每个智能体单独设定能读取的数据、能调用的工具和能执行的操作。Each agent gets its own scope: the data it can read, the tools it can call and the actions it can take.
- 最小权限原则Least privilege by default
- 高风险操作需人工审批（人在回路）Human approval for high-risk actions
- 运行过程可回放Every run can be replayed
合规就绪Compliance-ready
按国内监管要求设计交付流程，协助您完成上线前的合规工作。Delivery follows Chinese regulatory requirements, and we help with the compliance work before launch.
- 可在等保三级环境交付Delivery in MLPS Level 3 environments
- 协助生成式人工智能服务备案与登记Support for generative-AI filing and registration
- 生成内容按 GB 45438-2025 标识AI content labeled per GB 45438-2025
私有化部署Self-hosted deployment
数据不出域Data stays in your domain
等保三级环境交付MLPS Level 3 delivery
GB 45438-2025 内容标识GB 45438-2025 content labeling
备案与登记协助Filing & registration support
默认不使用客户数据训练模型No training on your data by default
SSO / RBAC
审计日志Audit logs
合作模式Engagement models
小步起步，
按价值扩展。Start small.
Expand on proven value.
可以从一次工作坊或一个 PoC 开始，验证价值后再扩大投入。每个阶段都有明确的交付物与通过标准。Begin with a workshop or a single PoC, and invest more once the value is proven. Every stage has clear deliverables and exit criteria.
探索Explore
联合创新工作坊Co-innovation workshop
1–5 天1–5 days
- 业务骨干与工程师同场工作Your experts and our engineers in one room
- 用真实数据现场搭建原型A working prototype on your real data
- 决定是否进入 PoCA clear go or no-go for a PoC
验证Validate
共创验证（PoC）Proof of concept (PoC)
2–4 周2–4 weeks
- 建立评测集与效果基线Eval set and quality baseline
- 验证技术可行性与数据就绪度Technical feasibility and data readiness
- 输出试点方案与投入估算Pilot plan and effort estimate
交付Deliver
项目制交付Project delivery
按里程碑与业务指标验收Accepted by milestone and business metric
- FDE 与您的团队联合交付FDEs and your team deliver as one
- 从试点上线到生产投产From pilot through production launch
- 移交源代码、评测集与运行手册Source code, eval sets and runbooks handed over
运营Operate
持续运营服务Managed operations
按约定的服务等级（SLA）提供Delivered under an agreed SLA
- 7×24 监控与分级响应24/7 monitoring and tiered response
- 模型升级与回滚Model upgrades and rollbacks
- 季度价值复盘Quarterly value reviews
周期为典型参考；具体范围、周期与费用以双方确认的工作说明书（SOW）为准。Durations are typical. Final scope, timeline and fees are set in a mutually agreed statement of work (SOW).
常见问题FAQ
您可能关心的问题。Questions we often hear.
FDE 与传统外包有什么区别？How is FDE different from outsourcing?
传统外包通常按需求说明书交付功能，验收即结项。FDE 以约定的业务指标为目标：工程师嵌入您的团队，用评测集和基线数据驱动迭代，上线后持续运营，并把源代码、评测集和运行手册移交给您。它不是人力外包，而是对结果负责的工程团队。
Outsourcing usually delivers features against a spec and ends at sign-off. FDE works toward agreed business metrics: engineers embed with your team, let eval sets and baseline data drive each iteration, keep operating after launch, and hand over the source code, eval sets and runbooks. It isn't staff augmentation — it's an engineering team accountable for results.
我们的数据会被用于训练模型吗？Will our data be used to train models?
默认不会。您的数据仅用于为您构建和运行系统；如需使用您的数据微调专属模型，我们会事先取得您的书面授权。采用私有化部署时，数据的存储、推理与流转均在您的环境内完成。
Not by default. Your data is used only to build and run your system, and we ask for your written approval before fine-tuning a dedicated model on it. With self-hosted deployment, storage, inference and data flows all stay inside your environment.
支持哪些大模型？Which models do you support?
架构与模型无关（Model-agnostic）。我们适配已备案的国产大模型与主流开源模型（如 DeepSeek、Qwen、GLM 系列），按每个场景的效果、时延与成本评测选型，并支持多模型路由。更换或升级模型时，应用层无需重写。
The architecture is model-agnostic. We work with domestic models that have completed generative-AI filing and with leading open-source models (such as the DeepSeek, Qwen and GLM families), choose by measured quality, latency and cost for each use case, and route across models where that helps. Swapping or upgrading a model doesn't mean rewriting the application.
一个场景从启动到上线需要多久？How long does one use case take to go live?
典型路径为：场景评估 1–2 周，共创验证（PoC）2–4 周，试点上线（MVP）6–12 周。影响周期的主要因素是数据与系统接口的就绪程度；场景评估结束时，我们会给出具体的计划与里程碑。
A typical path is 1–2 weeks to discover, 2–4 weeks to prove (PoC) and 6–12 weeks to pilot (MVP). The biggest variable is how ready your data and system interfaces are; you get a concrete plan and milestones at the end of discovery.
私有化部署需要什么硬件？What hardware does self-hosted deployment need?
取决于模型规模与并发量。7B–14B 级模型，通常一台 GPU 服务器即可支撑部门级应用；更大的模型或更高的并发需要多卡或多机部署。我们会根据评测与压测结果给出配置建议，所需算力也可以由 Branpont 智算基础设施提供；GPU 型号、数量与可用区域（境内 / 境外）以实际供应及出口管制政策为准。
It depends on model size and load. A 7B–14B model can usually serve a department-level application from a single GPU server; larger models or heavier traffic need multiple GPUs or nodes. We size the hardware from eval and load-test results, and the compute can also come from Branpont AI Infrastructure. GPU models, quantities and regions (mainland China or international) depend on actual supply and export-control policy.
如何衡量效果？How do you measure results?
启动时先记录基线、约定目标值，再以“基线 → 结果”的方式衡量，例如处理时长、问题解决率、准确率、人工投入或单次成本。技术指标（如评测集准确率、首词元时延）作为每次发布的门禁，业务指标用于验收与季度复盘。
We record a baseline and agree on targets before we start, then measure baseline against result — handling time, resolution rate, accuracy, manual effort or cost per task. Technical metrics such as eval-set accuracy and time to first token gate every release; business metrics drive acceptance and quarterly reviews.
从一个场景开始，
走向规模化。Start with one use case.
Scale from there.
告诉我们您希望改进的业务环节，解决方案架构师将与您联系，并建议从联合创新工作坊或 PoC 开始。Tell us which workflow you want to improve. A solutions architect will get in touch and suggest whether to start with a workshop or a PoC.
或直接发送邮件至Or email contact@branpont.com