Work order reader

Classifies maintenance work orders and finds the closest past cases. Built by Alvin Alias.

F1 94.4% / training 1.1% of params LoRA fine-tune · full fine-tune reference 95.9%
AA-2026-05
model waking, first run may take up to a minute

Rotates eight orders I wrote and labeled myself. The intended label prints after the model commits.

View or paste the work order

Paste your own

Waiting for a work order

0.00 confidence

your text · no intended label to compare

Every category score

Every category's score

Extracted fields

What the rule-based pass pulled out

regex extraction; the LLM extractor is a notebook result, not this endpoint (note 5)

No fields returned.
Similar past cases

Closest past cases from the corpus

No similar cases returned.
Calibration record · LoRA vs full fine-tune vs TF-IDF
Macro F1 94.4%DistilBERT + LoRA, 743K trainable params, 1.1% of the model, 2.8 MB artifact
95.9%full fine-tune, 67M params, 256 MB
93.5%TF-IDF + logistic regression baseline
Failure-mode extraction13% rule-based to 70% with the LLM extractor
Hybrid extractor56% fewer LLM calls but far less than 56% of the uplift kept; regex confidence is a poor escalation gate
Held-out test600 records, stratified, seed 42
How it works · three uses of language models

I use one pipeline three ways: extraction from messy text into structured records, a LoRA fine-tune that updates 1.1% of parameters, and semantic retrieval over past cases. The useful result is the negative one: the hybrid extractor's regex confidence gate chose the wrong records to escalate.

Notes · the corpus is synthetic and says so
  1. This ML demo sleeps after extended inactivity. First wake can take a moment; runs after that are quick.
  2. Every record in the corpus is synthetic and generated with a calibrated noise layer; the taxonomy, vocabulary, and abbreviations come from 12 years of writing real work orders at Rheem, Centurion, Baker Hughes, and Daikin.
  3. The eight samples are mine: written for this page and labeled with what I intended. When the model disagrees with me, the page says so.
  4. Similarity search retrieves from the training corpus; the returned cases are synthetic records.
  5. Field extraction on this page runs the rule-based pass, which is fast and free; the LLM extractor that lifts failure-mode accuracy from 13% to 70% is measured in the notebooks and priced in the calibration record.