PROFESSIONAL WORK / HUAWEI · NAIE
AI platform.
Model lifecycle engineering.
High-level professional work across NAIE/Federated service flows for optical-fiber sensing: dataset resolution, retraining and evaluation task assembly, online-learning device workflows, and versioned model-package delivery.
- Context
- Huawei · NAIE / Federated
- Contribution
- Java/Spring · API integration · Debugging
- Lifecycle
- Dataset · Train · Evaluate · Deploy
Connect the model lifecycle to the service layer.
The recent Huawei work I can publicly describe sits around NAIE/Federated workflows for optical-fiber sensing. The engineering surface is broader than one model call: service APIs, dataset metadata, retraining/evaluation tasks, online-learning devices, and model-package delivery all have to agree on identifiers and state.
Resolve the dataset
Map dataset identifiers to the storage paths expected by downstream tasks.
Assemble the task
Carry model version, datasets, metrics, and result locations into retraining or evaluation requests.
Deploy the artifact
Locate the versioned package, extract the model artifact, and dispatch it through the device workflow.
Data paths are part of the contract.
Dataset path resolution
Implemented and debugged dataset metadata handling around queryDatasetHofsPath, including conversion of returned collection data into dataset-to-path mappings used by training and evaluation flows.
Evaluation task assembly
Built Java request bodies carrying model version, evaluation result paths, dataset paths, high-quality dataset paths, and metrics such as Accuracy and Recall. When the serialized datasets field disagreed with the populated path configuration, I traced the mapping and serialization chain rather than treating the downstream API as a black box.
Service-specific request behavior
Investigated differences between interface conventions, including request headers, cookies, URL construction, and transport behavior. This helped separate an ER/IR integration mismatch from the business-level task logic.
Move a model package from repository to device.
The model deployment flow uses a known versioned HFS repository path. The service locates the package, extracts model.zip, and creates the corresponding online-learning deployment task for the target device.
I also worked with the surrounding package-dispatch and SFTP delivery logic, focusing on artifact identity, temporary extraction, task construction, and the boundary between cloud package handling and device-side deployment.
Infrastructure view: model deployment is a lifecycle problem, not just a file copy. Version identity, artifact selection, transport, task state, and device-side execution all form one chain.
Follow the failure across layers.
Check the wire contract
Inspect the actual URL, headers, cookies, and interface-specific conventions.
Check the data shape
Trace maps, DTO conversion, generic response types, and null initialization.
Check the serialized body
Compare the in-memory task model with the JSON finally sent to the service.
Examples included Spring URL parsing failures, HTTP 500 responses from deployment-information flows, empty dataset objects during evaluation-task creation, and object-mapping failures caused by an uninitialized converter. The useful lesson was to trace the request end-to-end: endpoint semantics, response shape, conversion, task construction, serialization, and downstream behavior.
What this case does — and does not — claim.
This page describes engineering work at a public-safe level. It does not expose internal URLs, credentials, dataset contents, proprietary package paths, customer information, or unpublished model metrics.
The work demonstrates practical AI platform engineering around model lifecycle orchestration and service integration. It should not be read as a claim of owning the entire NAIE platform or the underlying model algorithms.
Professional contribution summary. Internal implementation details are intentionally abstracted; no proprietary source or infrastructure identifiers are reproduced.