Machine learning for sensor and signal data
Your data holds more signal
than your model finds.
Artif Wonder is a machine learning company. We take the sensor and measurement data you already collect, find what it can predict, and deliver a working system that moves a number your business tracks. Simple, fast, and built to run where you need it.
Working software, not slide decks
The simplest solution that performs
Measured in your business numbers
Example deliverable
Feasibility audit
• GoSuccess metric and target
agreed
Physics-informed features
engineered
Baseline model on client data
trained
Achievable gain on the metric
measured
Model delivery
• Ready to deployRuns on
Your existing hardware
Features
Physics-informed
Predictions
Real time
Cloud dependency
None
Held-out data
• PassedField recording
• PassedOn-device test
• PassedHow we work
Start from the physics. Build the simplest model that works. Measure it in your numbers.
We start from the physics of your signal
Raw sensor data carries structure: frequencies, transients, phase, decay, dose response. We engineer features from that physics first, so the model learns from signal rather than noise and needs far less data to do it.
We build the simplest model that meets the target
A compact classifier where it reaches the number, a larger network only when the data earns it. Simple models run on your existing hardware, explain their decisions, and keep working after we leave.
We measure success in your metric, not ours
Accuracy is not an outcome. Before training starts we agree the business number the model must move: detection rate at a fixed false-alarm budget, hours of physical testing, cost per prediction, time to result. Every validation report is written against it.
Typical starting points
Years of recordings and no answer on what they can predict
Lookup tables and thresholds that stopped scaling
An existing model that misses events or drifts
A device that cannot afford the model it needs
How it works
Metric, features, baseline, verdict. Then build.
Every engagement starts with a fixed-price audit that ends in a measured answer on your metric. If you continue, the audit fee is credited toward the project.
01
2 to 4 weeks, fixed price
Feasibility audit
The gain your data can deliver, measured on your metric.
02
Milestone-based
Built
The simplest model that reaches the target, trained on your data under your NDA.
03
Handover
Deliver
Software that runs in your environment from day one.
What you get
Production software, built to your specification.
Designed to your constraints
We fix these with you in the audit. They decide the architecture.
Success metric
technical (detection rate, false alarms, latency, model size) or business (test hours, cost per unit, time to result)
Target hardware
your existing hardware, a server or the cloud
Latency
sub-millisecond to batch
Memory and power
kilobytes to gigabytes
Labels
none, few or many
Explainability
black box not fully auditable
You own what we built for you
✓ Source code and trained models built for your project
✓ Your NDA, signed before any data
✓ Work on your infrastructure if required
✓ Project data deleted on completion
Runs with what you already use
A Python package or a container your team calls from its existing tools. No new runtime to learn, no proprietary platform.
Python package
Docker container
REST API
Validated against your metric
Every delivery includes a test suite, integration documentation, and a validation report that states the result in the business metric agreed in the audit.
Test suite
Validation report
Integration guide
Selected work
Signal in. Business metric out.
RF sensing
Client project
30%
Improvement in signal detection in noisy environments for a consumer electronics product line
30% more signals detected in noisy environments, on the existing hardware. Rule-based detection had missed weak signals and confused them with interference. We combined physics-informed feature engineering with machine learning to sharpen frequency localization and separate true signals from noise.
Client
Consumer electronics manufacturer
Method
Physics-informed features on the raw RF stream; compact classifier on the existing hardware
Runs on
Portable devices and cloud test systems
Wearables
Client project
45%
Reduction in operating cost for real-time monitoring of physiological states
45% lower operating cost for real-time stress monitoring on a wearable. The inputs were continuous physiological signals, changing environments, low-power hardware and strict privacy. We decoded stress from skin conductance and heart rate with models that adapt from user feedback and explain their predictions.
Client
Digital health start-up
Method
Skin conductance and heart-rate features; compact adaptive model on the device
Runs on
Wearable device with cloud retrainin
Biotech instrumentation
Client project
35%
Enhancement in coil power for a high-precision magnetic field
35% less coil power at the same field precision. The client needed coils that produce a highly specialized magnetic field, at a cost that works for a product. We coupled a physics simulation of the field with reinforcement learning, which searched thousands of coil geometries and converged on designs that meet the field specification with the least power.
Client
Biotech company
Method
Physics-based field simulation; reinforcement learning to search the design space
Result
Field specification met with 35% less power; design fixed before the first prototype
Founding team – Est. 2019
Every engagement is led by the founders.
The founding team scopes, leads and delivers each project directly. No account managers, no hand-offs.
Marianna Yanike, Ph.D.
Co-Founder & CEO
Drives Artif Wonder’s business strategy, product development, and customer partnerships. Combines expertise in digital health, technology and biotechnology commercialization with hands-on experience bringing innovative solutions to market.
Ph.D. in Neural Science, NYU
LinkedIn profileArnulf Graf, Ph.D.
Co-Founder & CTO
Has led the technical delivery of every Artif Wonder client project since the company’s founding, across consumer electronics, biotech and digital health: scoping, model development and handover of working systems. Before that, built and shipped machine learning products in Silicon Valley.
Ph.D. in Machine Learning, Max Planck Institute · M.S. in Physics, EPFL
LinkedIn profileGet in touch with us to discuss your project.
Every project starts with a conversation about your data and the result you need. From there, we propose the simplest path to a working solution.
Contact to discuss