Core RAG & agent development
Organize retrieval, reasoning, tool calls and multi-step execution around business tasks. Connect documents, APIs and existing systems, with human confirmation for important actions and a record of execution.
AI AGENT & RAG ENGINEERING · HANGZHOU
Agents for
real work.
Hands-on RAG and agent engineering, backed by a 20-person team spanning full-stack development, algorithms, data and QA.
Written Japanese / Meetings in English
ENGINEERING, TOGETHER.
From knowledge to agent action.
01
KNOWLEDGE
Documents · Knowledge · Business data
02
AGENT CORE
Retrieval · Inference · Tool integration
03
WORKFLOW
APIs · Business systems · Deliverables
Evidence tracking · Evaluation · Testing
From core logic to business integration.
20People available for sustained delivery
4Connected engineering disciplines
2Project delivery / Full-time engagement
01 / AGENT ENGINEERING
From core logic
to business integration.
Team members bring hands-on experience implementing core RAG and AI-agent logic, and independently developing LLM tools for exam authoring and review.
MEMBERS’ HANDS-ON EXPERIENCEKnowledge & evidence systems
Team members bring experience with knowledge bases, knowledge graphs and exam-authoring tools. We can design document structure, source metadata and citation detail so users can return to the original evidence.
Multi-model integration & operations
Connect model providers through authentication, routing, tenant isolation and usage records. Match models and context to the task, evaluating answer quality, response time and operating cost.
RAG / KNOWLEDGE ENGINEERING
Scattered documents.
Traceable answers.
Retrieval-augmented generation (RAG) finds relevant evidence in your information before a model composes an answer. We extend knowledge-base development into document governance, access control and evaluation, so the system can be checked and kept current.
- 01Business sources→
- 02Parse & index→
- 03Retrieve & rerank→
- 04Answer & cite→
Connect and organize sources
Scope PDFs, Word documents, spreadsheets, web pages or business data. Define categories, owners and update rules, and establish which sources may be used in answers.
Parse and structure documents
Handle OCR, layouts, headings and tables. Split content by meaning and business structure while preserving filenames, pages, sections and versions for source tracing.
Combine retrieval methods
Use keyword and vector retrieval, with query rewriting, metadata filters and reranking where needed. Test representative questions to find missing or irrelevant evidence.
Generate with evidence
Build answers from retrieved content with citations that locate the source. Define how to handle insufficient evidence, outdated documents and version conflicts, including clarification or human review.
Control access and updates
Filter retrieval by user, department or tenant. Separate upload, approval and publication rights, and agree on addition, update, removal and index synchronization workflows.
Evaluate and improve
Create an evaluation set of questions, reference answers and source evidence. Review retrieval, citations, answer quality, latency and cost, then iterate using feedback and failure cases.
Illustrative knowledge Q&A · Sample content
Who approves this purchase, and which documents are required?
Retrieve the workflow for the purchase type and amount, with references to the applicable policy. If the amount is missing or multiple versions exist, clarify those details before giving guidance.
Sample references
Purchasing policy · Approval workflow · Section 4
Purchase request guide · Required documents · Section 2
Help users understand the evidence, the applicable version and what to do next.
Possible deliverables
- Source ingestion and parsing pipeline
- Knowledge index and retrieval / Q&A APIs
- Citations and access configuration
- Evaluation set, failure cases and improvements
- Deployment guide and operational handover
Scope is agreed around source formats, data permissions, existing systems and acceptance criteria.
Discuss a knowledge-base projectRECENT PRODUCT / WENQU EQ-AGENT
WenquEQ-AGENT
An agent for the exam-authoring workflow.
A recently developed, local-first agent product connecting source reading, question generation, review and output archiving. It integrates an open-source agent core with a domain-specific workflow.
- 01Understand
- 02Read sources
- 03Generate
- 04Review & revise
- 05Archive
ENGINEERING FOCUS
Tool execution & multi-agent coordination
File operations and validation steps, with coordinated planning and execution.
Reusable Skills
Structured expert workflows, quality standards and checklists for repeatable use.
Context, permissions & audit trails
Long-running task context, scoped permissions, approvals and execution records.
Local execution & intranet models
Workflows running in a local application, connected to intranet-hosted models.
The current case shows high-school mathematics question generation, self-checks and Markdown output.
02 / ENGINEERING CAPABILITIES
The systems
around your agent.
We define the scope and technical approach around your goals and existing environment.
RAG & knowledge systems
- Ingest business sources
- Retrieve and rerank
- Answer with citations
Link answers to the source location and version for review.
Build searchable, traceable and maintainable knowledge bases for policies, product documentation, technical material and business questions, connected to your workflow.
- Development scope
- Source ingestion, parsing and chunking, vector and keyword retrieval, reranking and citations; knowledge graphs and business relationships where required.
- Integration
- Knowledge Q&A APIs, internal search interfaces or agent tools, with knowledge spaces organized around departmental and tenant permissions.
- Delivery & validation
- Validate retrieval and answers against representative questions. Hand over indexing pipelines, APIs, evaluation results and maintenance instructions.
OCR & document processing
- Scans and PDFs
- Extract and review
- Store structured data
Route low-confidence fields to a human reviewer.
Turn images, scanned documents and PDFs into usable business data, connecting human review with downstream processing to reduce repeated data entry.
- Development scope
- OCR, layout and table parsing, field extraction, format normalization and business-rule validation.
- Processing workflow
- Batch files or API ingestion, low-confidence field flags, and optional human review, exception handling and processing-status tracking.
- Delivery & validation
- Structured outputs, processing APIs and field mappings, evaluated for field accuracy and processing time using customer samples.
Model inference & deployment
- Connect models
- Deploy inference
- Serve business apps
Validate latency, throughput and resource cost at agreed loads.
Design inference environments around models, data boundaries and concurrency needs. Validate performance, stability and operating cost against available hardware and budget.
- Development scope
- Local or private deployment, model-service connections, unified calling APIs and access authentication.
- Operational tuning
- Evaluate quantization, batching, caching and model routing as required. Record latency, throughput, resource use and errors.
- Delivery & validation
- Deployment configuration, performance and cost comparisons, monitoring plans and upgrade / rollback instructions, tested at agreed load levels.
Full-stack & data engineering
- Web workspace
- APIs and workflows
- Data and integrations
Check access controls and regressions across the UI, APIs and data flows.
Bring model capabilities into the systems people use every day, covering frontend, backend, data processing and testing.
- Application development
- Web workspaces, administration interfaces, APIs, roles and workflows, integrated with existing systems and third-party services.
- Data engineering
- Data ingestion, cleaning, transformation, scheduled jobs and data-quality checks, connecting analytics and business applications.
- Quality & handover
- Functional, API and integration testing, with source code, API documentation, deployment instructions and operational handover materials.
Delivery support by project stage
For Japan-facing collaborations and existing systems, agree on scope, deliverables and customer acceptance for each stage.
Requirements and acceptance design
Map business flows, system boundaries and exceptions. Align specifications, changes and acceptance criteria through written Japanese documentation.
- Example deliverables
- Requirements and interface notes, a change log and acceptance test cases.
Existing system integration and phased migration
Map dependencies and data structures, connect through APIs or batch jobs as appropriate, and plan data reconciliation and cutover rehearsals.
- Example deliverables
- A dependency inventory, field mappings and cutover and rollback procedures.
Test automation and release handover
Set up API and regression tests within the agreed scope and add key checks to CI. Define how access, audit logs, releases and alerts will be handed over.
- Example deliverables
- Test results, a release checklist and an operations handover guide.
03 / THE TEAM
Specialist skills.
A connected team.
A stable capacity of 20 people. We assemble the right mix of roles for your project’s scope and stage.
Our team includes engineers with experience at Alibaba DAMO Academy’s Artificial Intelligence Center.
20people
ENGINEERS & QA
10people
Full-stack development
Web · APIs · Integration
5people
AI & algorithms
Models · Retrieval · Inference
2people
Data engineering
Processing · Pipelines
3people
Testing & QA
Functional testing · Quality checks
Different disciplines. Shared delivery.
Illustrative development scenes; these are not photographs of our team or office.
Engineering collaboration
Align APIs and technical boundaries across full-stack, AI and data roles.
Compagnons / Unsplash ↗
Requirements & design
Turn business needs into scope, acceptance criteria and an iteration plan.
Creatopy / Unsplash ↗
Implementation & validation
Connect development and QA around code, interfaces and running results.
Bernd Dittrich / Unsplash ↗04 / WORKING TOGETHER
The right shape of collaboration.
Match the engagement to your scope, specialist needs and timeline.
PROJECT DELIVERYA
A team for your project
We assemble a team to deliver an agreed scope, from development and validation through to handover.
- Defined scope and deliverables
- Progress shared against milestones
- Agreed acceptance criteria
FULL-TIME ENGAGEMENTB
Capacity for your team
Engineers join your development workflow for an agreed duration, with roles and capacity matched to your needs.
- People matched to the required roles
- Collaboration with your existing team
- Agreed capacity and duration
Written Japanese. Meetings in English.
Align requirements and decisions in writing, with technical discussions in English.
05 / FROM BRIEF TO DELIVERY
A clear path to handover.
- 01
Understand the brief
Clarify workflows, challenges, priorities and your existing environment.
- 02
Agree on scope and team
Define deliverables, responsibilities, milestones and validation methods.
- 03
Build and validate
Share progress and develop in stages, with testing throughout.
- 04
Deliver and hand over
Review agreed outputs and transfer the information needed to operate them.
LET’S BUILD SOMETHING USEFUL
Tell us what
you’re building.
Your requirements, preferred timeline and the people you need. Let’s start with what you know.
kxbkey@gmail.com
The button opens your email application.
- Company
- 杭州云眼数据科技有限公司
- Location
- Hangzhou, Zhejiang, China