02SolutionsCases

Four product lines,
four real scenarios

All cases below are projects we have delivered or are actively developing. Client names are anonymized to protect client information.

Delivered: 03
In development: 01

AI Smart Classroom · On-premise agent "Xiaozhi"

An on-premise AI teaching assistant for training classrooms, addressing three core needs: the lecturer is never interrupted, questions are handled under control, and classroom content is preserved.

Before class · Build the course knowledge base
Lecturer courseware (PPT / PDF / Word) is parsed multimodally; text, images, charts and page structure become course knowledge in a per-course knowledge base.
During class · Question parking lot
Learners ask on desk tablets; questions flow into the lecturer's private screen, auto-clustered, de-duplicated and ranked by heat. The lecturer decides to answer now, at the end of the section, or after class. The system retrieves course knowledge to draft candidate answers with cited evidence, reviewed before output.
After class · Preserve organizational knowledge
A one-page summary, topic-clustered FAQ and source list are generated automatically and preserved as reusable, auditable organizational knowledge.

Key capabilities

Multimodal course understandingContent recognition · image-text parsing · semantics
Knowledge constructionChunking · semantic organization · vector indexing
RAGQuery understanding · precise retrieval · content recall
Local generationContext fusion · answer composition · wording refinement
Trusted outputCourse constraints · content review · safety controls
Two-tier knowledgePer-class library + global library with review loop

Illegal Dumping Vehicle Monitoring Platform

Intelligent recognition and evidence capture of illegal dumping vehicles across a township's camera network, compressing "spot a suspicious vehicle" to "evidence retained" into a single human confirmation.

Recognition
Load state (empty / loaded) of every vehicle entering camera view is recognized in real time, with snapshots and detection boxes recorded.
Matching
Empty vehicles are compared against loaded vehicles seen at nearby cameras within the previous 40 minutes; matches above threshold are pushed to human review, and high-confidence events are pinned and counted in the alert bar.
Review
Side-by-side view with snapshots and historical playback, zoom into the cargo bed, one-click "abnormal / normal" judgment with automatic advance to the next item.
Evidence
Confirmed vehicles are retained automatically and a suspected route between camera sites is drawn on the map to assist enforcement.

Platform components

MapCamera sites · status colors · alert ripple · vector / satellite toggle
Live videoLive streams and playback from the video cloud platform
PanelsVehicle search · pending review · processed (normal / abnormal)
DashboardDaily / weekly / monthly recognized, auto-reviewed, human-reviewed and confirmed dumping
AccountsAdmin-provisioned users · session expiry · subscription control
Auto refreshPending items and stats every ~10 s, online status every ~5 s

Office Automation (OA) System for a Technical College

A dual-client collaboration system for college staff, centered on configurable approval workflows and notifications delivered to WeChat.

Workflows
Leave, document circulation with multi-branch routing, meeting rooms, official vehicles, vehicle maintenance, seal usage, bus rental and maintenance requests, built on the Flowable engine.
WeChat integration
WeChat login and account binding, service-account template notifications, SMS activation and password recovery.
Administration
Organizations, users, roles, workflow templates, document numbering, attachment upload and download.
Mini-program
To-do / done / my requests, submit and approve, process details, document lookup.

Delivery & operations

BackendJava · Spring Boot · MyBatis · Flowable · Redis · RabbitMQ
FrontendVue · Element UI · native WeChat mini-program (TypeScript)
DeploymentCloud server · Nginx · systemd · Docker middleware
EnvironmentsProduction and staging, full handover docs and schema dictionary

Theory–Simulation Agent Platform

For simulation researchers: "MATLAB theoretical calculation → mapping table → PSCAD simulation → templated Word report" fixed as a traceable workflow, with an agent that assists rather than replaces the researcher's decisions.

Workbench
Parameter table, four-step pipeline bar, theory curves vs. simulation waveforms, run logs; the full pipeline runs from buttons even without an LLM.
Agent
The chat panel calls only registered tools (MCP); no improvised scripts, no skipped steps; theory and parameter suggestions first, simulation only after user confirmation.
Connectors
MATLAB via Engine or official CLI; PSCAD via Automation or a CLI fallback; mapping tables constrain writable parameters.
Delivery
Windows desktop application; all computation stays on the local machine; report text comes from results, not from the model.

Design principles

Separation of concernsConnectors do atomic work, workflows own ordering and checkpoints, agents plan and call
ReproducibleParameter changes invalidate downstream checkpoints
No masqueradingProduction paths fail loudly; no mock results stand in for real ones
No silent startupDetecting installed software is separate from launching it

Want something similar
in your environment?

Tell us about the business, the environment and the boundaries, and we will start from requirements together.