swiftclf-tuna: bilingual hierarchical intent classification for assistant routing
From a single user message the model determines the request type and picks a route: retrieval, generation, tools, agent execution, clarification or fallback. The taxonomy has 4 top-level classes and 22 second-level ones. Selective prediction with a fallback branch and risk-aware metrics.
Methodology. AI-orchestrated research: problem framing, taxonomy and evaluation criteria design, direction of an autonomous LLM-agent workflow and human-in-the-loop validation of results — the founder; implementation, experiments and documentation were done by an agent workflow with counter-review.
Metrics are from the current internal evaluation report of the model on the holdout set; the project is not presented as SOTA or an academic publication. Artifacts include model cards, the training and evaluation pipeline, and reports.