About the implementation and examples
Engineering notes draw on author-supplied MailTrace development reports through 13 October 2026. They describe the documented revision, including local work, rather than independently verified live deployment. Experiment designs and architecture proposals describe work to evaluate. Examples are educational, use synthetic data and are not copied production code.
Separate the model from the framework and the product
A framework provides tools for defining and running learning computations. A pretrained model supplies learned parameters. The product still needs ingestion, permissions, source storage, review and a usable interface. Adding a framework does not automatically add any of those product capabilities.
Possible future experiments include adapting a document classifier or evaluating a local model for a tightly bounded task. The supplied records do not establish a TensorFlow or PyTorch training pipeline in MailTrace. They do describe external embeddings and LLM integrations, which should not be relabelled as in-house deep-learning training.
Choose ownership based on a concrete requirement
This decision outline is an architecture aid. It avoids treating hosted and local inference as universally better alternatives; each has operational and data-handling consequences that must be evaluated.
{
"hosted_pretrained": ["provider integration", "bounded inputs", "usage tracking"],
"local_pretrained": ["runtime packaging", "hardware budget", "model updates"],
"custom_training": ["reviewed labels", "training pipeline", "held-out evaluation"],
"shared_requirements": ["source permissions", "provenance", "review", "rollback"]
}Keep the experiment outside the evidence core
If a framework experiment is justified, give it a narrow input and output contract. It should consume authorised prepared data and return a versioned proposal. It should not create a parallel case database or bypass the existing review workflow.
Before adoption, compare quality, package size, runtime memory, operational cost and maintenance effort against the existing approach. For Android, local inference also needs device-specific testing and a clear offline use case. The professional choice is the smallest maintained system that meets the measured requirement, not the largest collection of frameworks represented in a course syllabus.
References & further reading
Official documentation for the methods discussed. Application-specific details come from the development records described in the series note.