Project demonstrations
These excerpts from real projects trace the work from model development and Agent tool use to engineering workflows and industrial deployment. Redaction is limited to account details, contact information, and business-sensitive material. Some clips come from the author's public Bilibili channel and have been trimmed and muted for the web.
EvoAI Vision Model Workflow
An instance-segmentation workflow spanning dataset management, label ontology, annotation, training, and model artifacts.
Enterprise Knowledge Agent
An end-to-end task in which an agent moves from data analysis to a finished document through a sequence of tool calls.
Data Analysis Agent
Reads multiple spreadsheets and analysis instructions, then uses tools to produce scripts, charts, and a consolidated report.
Document Understanding & Structured Capture
Reads printed, handwritten, and traditional-Chinese documents and turns key information into editable structured fields.
3D Assembly Copilot
Natural-language requests locate parts in an assembly and become visible operations in the 3D workspace.
High-speed Visual Sorting
Real-time industrial vision recognition and sorting on a high-throughput line.
Visual Privacy Redaction
Detects people and license plates in video and produces a privacy-safe deliverable automatically.
WebSeg Interactive Annotation
A lightweight web workspace for object detection and segmentation annotation, maintained as an open-source project.
Web-based Robot Control
Controls a quadruped robot from a tablet web interface while streaming camera footage and device state.
Enterprise Agent Platform
One governed agent runtime connecting models, knowledge, and tools so teams can complete real work.
- 12systems integrated
- 10reusable Skills
- 99.93%availability
- Challenge
- AI pilots across an enterprise often rely on different knowledge sources, tools, and controls. The challenge was to bring them into a shared platform that could understand tasks, invoke enterprise capabilities, complete work, and improve through evaluation.
- Role
- Platform and R&D lead for a 50+ person organization spanning frontend, backend, algorithms, data, testing, and operations.
- Key decisions
- Treat models, knowledge, data, tools, Skills, permissions, audit, evaluation, and operations as one runtime rather than a collection of independent features.
- System
- Task planning, MCP and connectors, Skill lifecycle, sandboxed execution, traceable citations, human confirmation, risk controls, regression evaluation, and observability.
- Outcomes
- Integrated 12 internal systems and delivered 10 reusable Skills. Acceptance evidence included 2.3-second average standard QA, 20-concurrency stability over 72 hours, 99.93% availability, and 95% retrieval accuracy.
Knowledge Agent →Data Analysis Agent →Document Intelligence →
Industrial Multimodal and Engineering Agents
Drawings, 3D models, BOMs, and process knowledge in one inspectable engineering workflow.
- Drawings2D understanding
- STEP3D feature evidence
- Process & quoteinspectable output
- Challenge
- Engineering work spans drawings, 3D models, BOMs, process knowledge, costing, and specialist tools. The system has to understand those objects and produce results that fit the actual workflow.
- Role
- Product, platform, and R&D lead connecting multimodal models with engineering workflows and delivery constraints.
- Key decisions
- Represent CAD/STEP evidence, BOM relationships, processing features, similarity retrieval, manufacturability, and quotations as inspectable intermediate results, so engineers can review how a recommendation was reached.
- System
- Drawing and 3D understanding, similar-part retrieval, process generation, automated quotation, human review, and feedback capture.
- Outcomes
- A reusable path for engineering agents: understand a part, retrieve similar work, propose a process, and generate a quotation draft, with evidence available at every step.
Model and Data Loop
A reproducible model lifecycle connecting data, training, evaluation, deployment, and production feedback.
- 120+datasets
- 2M+labeled images
- 99.5%recognition accuracy
- Challenge
- Continue improving a multimodal recognition system while preserving the ability to reproduce every data, training, evaluation, and deployment step.
- Role
- Software and algorithm lead for the model system, cloud inference, annotation and training platforms, and engineering delivery.
- Key decisions
- Build the lifecycle around data versions, labeling, experiments, evaluation baselines, deployment artifacts, and online feedback.
- System
- 120+ datasets, more than 2 million labeled images, 100+ model outputs, and a continuous path from collection to production feedback.
- Outcomes
- Recognition accuracy moved from 92% to 99.5%. Inference speed increased by 50%.
EvoAI Vision Model Workbench
A privately deployable workbench for dataset import, annotation, training, testing, and inference delivery.
- 4 task typesclassify, detect, segment, anomaly
- ONNXstandard delivery
- Privatedeployment boundary
- Challenge
- Industrial vision projects depended on scattered scripts and manual handoffs, leaving data, training, and delivery disconnected. The goal was one system that could accept raw data and produce a trained, testable, deliverable model.
- Role
- End-to-end design from product workflow and system architecture to the training runtime and delivery boundary.
- Key decisions
- Use the project and its live dataset as the unit for data state, annotations, training jobs, model artifacts, and test feedback. Make large imports and training recoverable asynchronous jobs.
- System
- Classification, detection, segmentation, and anomaly detection, with dataset import, image annotation, training jobs, model registry, single-image testing, feedback capture, and ONNX/Windows inference delivery.
- Outcomes
- A privately deployable vision workbench that keeps data, training, models, and deployment artifacts in one traceable workflow.
Reinforcement Learning, Robotics & Real-time AI
Connecting distributed policy training, web-based robot control, and visual inspection into real-time perception, decision, and feedback systems.
- 1,500+CPU cores
- 300concurrent devices
- 99%detection accuracy
- Challenge
- Complex online environments require decision policies, robot control, visual perception, and device feedback to work together while preserving clear safety boundaries.
- Role
- Algorithm researcher for environments, distributed learning, visual perception, evaluation, and production deployment.
- Key decisions
- Separate environment fidelity, policy learning, UI perception, and online evaluation. Design the anomaly pipeline around measurable operational acceptance rather than a single accuracy figure.
- System
- A distributed RL platform using 1,500+ CPU cores; web control, live trajectories, and camera streaming prototypes for AGVs and quadruped robots; and a visual anomaly detection service supporting 300 concurrent devices.
- Outcomes
- The UI-driven game agent reached Diamond 9. Robot control and live feedback were validated end to end. Visual anomaly detection achieved 99% accuracy and reduced detection time to 30 minutes.
Watch the robot demo →AGV navigation originalLive trajectory originalRL algorithm notes ZH →
Longitudinal Personal AI Platform
Context that persists over time, tools that write real records within policy, and human confirmation for consequential actions.
- Temporal memoryfact versions & forgetting
- Role isolationidentity & data boundaries
- Human reviewcontrolled actions
- Challenge
- A personal AI needs to understand facts, preferences, goals, and real-world activity as they change over time, while preserving strict boundaries between identities, roles, and data.
- Role
- Architecture and ongoing evolution of the agent runtime, domain model, data and memory systems, AI toolchain, and production delivery foundation.
- Key decisions
- Model chat history, current facts, business records, knowledge evidence, and temporary execution state separately; isolate long-term memory by identity and subject; version changing facts, make forgetting auditable, and retain human confirmation for consequential judgments.
- System
- Streaming agents, typed tool calls, temporal long-term memory, vector recall, evidence retrieval, multimodal inputs, longitudinal records, multi-role permissions, audit, idempotency, and safety policy.
- Outcomes
- Conversation can continue across time, tools can read and write real records within policy, and AI advice, deterministic rules, and human decisions remain explicitly separated.
- Identity-scoped
- Evidence-traceable
- AI/rule separation
- Human confirmation
For conversations about AI team building, agent platforms, or production deployment, contact xujiayu0606@gmail.com.