Business Context
AspenTech wants to expand an internal industrial AI copilot used by engineers across AspenTech Inmation and Aspen Mtell workflows. The first production need is to classify incoming user requests so they can be routed to the right LLM workflow: direct Q&A, retrieval-augmented generation over product documentation, structured extraction from incident notes, or agentic multi-step troubleshooting.
Data
You have 180,000 historical support and engineering requests collected from AspenTech products over 18 months.
- Task: 4-way text classification for workflow routing
- Labels:
qa (38%), rag_doc_lookup (31%), structured_extraction (17%), agentic_troubleshooting (14%)
- Text length: 20-900 tokens, median 140
- Language: English only
- Input sources: support tickets, operator notes, maintenance logs, and chat transcripts
- Noise: product codes, tag-like metadata, copied stack traces, sensor IDs, and duplicated boilerplate
Success Criteria
A good solution should achieve macro-F1 >= 0.84, recall >= 0.90 on agentic_troubleshooting, and p95 inference latency < 120 ms per request in batch-online serving. The model should be robust to domain terminology such as alarms, historians, asset models, work orders, and process deviations.
Constraints
- Must run inside AspenTech-managed infrastructure
- Fine-tuning budget is limited to a single A10/T4-class GPU
- Explanations for routing decisions should be inspectable by product and support teams
- Misrouting agentic cases is more costly than over-routing simple Q&A
Requirements
- Build an NLP pipeline to preprocess noisy industrial text and train a routing classifier.
- Use modern Python tooling and a transformer-based implementation.
- Explain how your design supports LLM, generative AI, and agentic AI workflows.
- Include evaluation, thresholding, and error analysis for high-cost misroutes.
- Describe how you would extend the classifier with retrieval or entity extraction features if performance plateaus.
