swiftclf-tuna
Bilingual (RU/EN) hierarchical intent classification for routing AI assistants. AI-orchestrated research: selective prediction, calibration, CPU-first deployment.
SOURCE: MANUAL README
swiftclf-tuna is a bilingual (RU/EN) hierarchical intent classification system for routing AI assistants. From a single user message, the model determines the request type and picks a route: retrieval, generation, tools, agent execution, clarification, or fallback.
Methodology: AI-orchestrated research
The project is not positioned as standalone academic research done by hand. My role is problem framing, taxonomy and evaluation-criteria design, directing an autonomous LLM-agent workflow, reviewing key decisions, and human-in-the-loop validation of results. Implementation, experiments, and documentation were carried out by an agentic workflow with planning, counter-review, and iterative verification.
Current metrics
| Metric | Value |
|---|---|
| Taxonomy | 4 L1 / 22 L2 classes |
| Holdout set | 2,625 examples |
| Coverage | 85.14% |
| Accepted accuracy | 87.07% |
| Top-1 / Top-2 accuracy | 80.88% / 89.68% |
| Coverage @ Risk ≤ 0.14 | 87.24% |
| Expected Calibration Error | 0.061 |
| p95 latency (1 CPU core) | 77 ms |
| Model size | 278M parameters |
Selective prediction with a fallback branch and risk-aware metrics; reproducible evaluation on a holdout; CPU-first deployment. Artifacts include model cards, training / eval pipelines, and holdout reports.
The metrics reflect the current verifiable result of an agent-assisted workflow. The project is not presented as a SOTA claim or a finished academic publication.
Public repository: github.com/ascorblack/swiftclf-tuna-research