Nokia

I designed, trained and evaluated AI models in Python with TensorFlow and PyTorch over Nokia network data in Budapest, working on the NLP side. The problem was turning written intent statements, the plain-language descriptions of what a network is supposed to be doing, into structured data the network tooling could act on directly, instead of an engineer translating each one by hand, along with the data pipelines that fed them. I worked with feedback from the research engineers on the team, iterating in short cycles so it became clear early which approaches were worth continuing.

Semantic parsing needs paired data: the sentence, and the structure it is supposed to produce. That pairing does not exist until somebody makes it, so a good part of the work was building the corpus and annotating it, which meant deciding what fields an intent statement even has before anything could be trained to fill them.

Before anything was trained I built the evaluation set and the scoring harness. A parse gets scored two ways: exact match on the whole structure, and slot-level F1 on the individual fields, because a parse can be ninety per cent right and still unusable. The hard cases were the ambiguous ones. Sometimes a sentence has two valid readings and the model picks the wrong one, but the fault is not really the model's. The intent was never written precisely enough to have a single answer.