Edge AI
Machine Learning in 21 KB: Our WebAssembly Kernel
Kutup moved its client-side machine learning into a 21 KB WebAssembly kernel written in C — no TensorFlow.js, no 50 MB model download. How it works and what it buys.
Browser analytics with C++ and WebAssembly
Compute locally with a small numerical kernel, explicit memory limits and reproducible checks.
In short
- WebAssembly runs compiled C++ functions in the browser, but for a small dataset the transfer overhead may exceed the computation savings.
- Typed arrays should be passed with explicit lengths, input size should be limited, and non-finite numbers should be rejected.
- Native results should be checked against an independent reference with a stated tolerance, and zero variance should be handled explicitly.
Choose the computation boundary
WebAssembly can run compiled C++ functions inside a browser. Kutup Lab uses it for statistics and simple prediction. Parsing, interface state and chart rendering remain in JavaScript. For a small dataset, transfer overhead may exceed computation savings; measure on actual devices before making speed claims.
Specify memory ownership
Pass typed arrays with explicit lengths. Limit input size, reject non-finite numbers and copy results before reusing memory. Never pass a length exceeding the buffer capacity. A fixed upper bound makes the behavior of an educational tool easier to reason about.
Check against an independent calculation
Use fixtures containing a constant series, one sample, negatives, a large offset and an outlier. Compare native results with a reference implementation using a stated tolerance. Handle zero variance explicitly instead of emitting infinity. Show which engine was used when WebAssembly cannot load.
Explain the privacy boundary
Local computation does not make an entire application offline. Lab data stays in the browser, whereas NQ questions go to its answer service. Explain this at the point of data entry. WebAssembly isolation does not replace application-level validation.
Frequently asked
How do I check whether WebAssembly actually speeds up browser analytics?
Measure on actual devices, since for a small dataset the transfer overhead may exceed the computation savings. Check native results against an independent reference implementation with a stated tolerance. Use fixtures with a constant series, one sample, negatives and an outlier, and show which engine ran if WebAssembly cannot load.
Does running analytics locally in the browser keep all data private?
Local computation does not make an entire application offline. Lab data stays in the browser, but NQ questions are sent to its answer service. Explain this at the point of data entry, because WebAssembly isolation does not replace application-level validation.
Sources and further reading
Pilot steps in this guide are proposed engineering workflows. Examples do not constitute measured field performance.
Questions on this topic
Keep reading.
Full library
Reliability
Remaining useful life (RUL): report an interval, not one number
Learn how RUL is defined, what each of three common approaches needs as data, and how to report P10, P50 and P90 values with a worked trend calculation.
ML
Isolation Forest for sensor anomalies: how it works and where it misleads
Read flagged rows in a sensor table by their isolation score, understand what contamination does in the model, and see where the method misleads you.