Business news
Matched intended label“The company reported higher earnings and raised its forecast.”
- Model output
- business
- Intended label
- business
Constant Systems / Local text intelligence
Small models.
Useful labels.
Your hardware.
Turn text into a label inside your own application. We help you build a focused classifier, integrate it, and measure how well it works on your task.
Local inference. No model API call required.
RECORDED LOCAL RUN / NEWS TOPIC
“The company reported higher earnings and raised its forecast.”
Model output
business
533,464-byte model artifact. Local CPU inference. This is a saved result from September 14, 2026.
01 / The model
A classifier for a defined set of labels, with its tokenizer and dependencies.
02 / The integration
A working path from your application’s text to the model’s output.
03 / The evidence
A comparison on your examples and hardware, with errors and limitations documented.
Inspect the work
Real outputs. Including the mistake.
These six authored examples were run locally on a CPU. They demonstrate working capability; they are not a representative accuracy test or live browser inference.
Sports news
Matched intended label“The football team won the championship after scoring two late goals.”
- Model output
- sports
- Intended label
- sports
An ordinary text message
Matched intended label“Are we still meeting for lunch tomorrow?”
- Model output
- ham
- Intended label
- ham
An assistant command
Matched intended label“Play some jazz music”
- Model output
- PlayMusic
- Intended label
- PlayMusic
A movie review
Matched intended label“An entertaining and delightful film.”
- Model output
- positive
- Intended label
- positive
A tricky movie review
Known mistake“The acting was not bad.”
- Model output
- negative
- Intended label
- positive
The model missed the negation in “not bad.” A high model score does not guarantee a correct answer.
AURE runtime / controlled local comparison
Less time for the same checked labels.
0.1234 ms
Original runtime
0.0930 ms
Optimized runtime
24.6%
Lower measured latency
Median of process p50 latencies across four fresh-process runs per configuration; 3,072 timed calls per configuration using 256 sentences. Includes tokenization, remapping, inference, softmax, and result construction. Excludes startup, network, and queueing. All 872 validation labels matched the original model; unchanged predictions do not establish correctness. This compares our original and optimized AURE paths, not a competing product. Your workload needs its own comparison.
Read the measurement summary →Start with one task
Find out where a small model fits.
A good first conversation starts with your text, the labels you need, the device you run, and what a wrong answer costs. News categorization, SMS spam classification, and a fixed set of assistant commands are demonstrated starting points.
Customer-specific labels, training, integrations, and deployment are scoped separately. We agree on acceptance criteria and evaluate against your current approach before claiming an advantage.
Start with a free discovery conversation →Existing consulting offer
System Audit
$500 USD / one time
A review of your AI or infrastructure with a written report and a 30-minute follow-up. Use it to assess where local classification may fit.
- System architecture review
- Security and performance assessment
- Written recommendations
- 30-minute follow-up call
Model training and a custom build are separate engagements. No accuracy or savings guarantee is included.