Industry · Biomedical
Biomedical systems
A wearable biosensor has to be most accurate exactly where it is hardest: at low signal, under rapid change, inside a shrinking power and size budget. Agentriq treats that trade-off as an engineering problem you can simulate, learn from, and close a loop on.
01The problem
Continuous glucose monitors from makers such as Abbott and Dexcom are least accurate in the low-glucose, rapid-change regime where a wrong reading is most dangerous. Published point-accuracy studies show mean absolute relative difference roughly doubling below 70 mg/dL, and worse still when glucose moves faster than 3 mg/dL per minute, while power, latency and footprint all have to keep shrinking. Accuracy at the edge of the signal, not average accuracy, is the figure that decides safety.
02How we solve it
Simulate the device, learn the signal.
Agentriq couples physics-based device simulation with deep learning on physiological signals, then optimizes both against the low-signal regime that actually fails.
Simulation-driven biosensor and microfluidic design
Multiphysics models in COMSOL and ANSYS Fluent couple diffusion, electrochemistry and flow to predict sensor response and refine electrode geometry before any device is fabricated.
Deep learning for physiological signals
Convolutional and recurrent networks trained on MIT-BIH and PhysioNet records reach roughly 99 percent arrhythmia classification accuracy on ECG, with matching pipelines for EEG and PPG.
Multi-objective device optimization
Accuracy, power, latency and footprint are optimized jointly, so gains in the low-glucose regime are not silently paid for in battery life or size.
Edge and FPGA real-time inference
Quantized, pruned models are mapped to FPGAs with Xilinx Vitis AI for sub-watt, millisecond-scale inference running on the wearable itself.
03What it produces
Sub-watt inference at full accuracy.
In published edge deployments, a compact FPGA runs per-beat ECG arrhythmia inference in roughly 11 milliseconds at about 0.33 watts while holding near 99 percent accuracy, the kind of accuracy, power and latency budget a wearable can actually meet.
04The agentic loop
Every design run verified and judged.
Agentriq runs the whole pipeline as a closed agent loop: it proposes a design, simulates and scores it, verifies the result, finds the root cause when a target is missed, and writes the decision back.
- 01Propose a device geometry, physiological-signal model and edge configuration from the current targets and constraints.
- 02Run the multiphysics simulation and signal-model training, then score accuracy, power, latency and footprint against the low-signal regime.
- 03Verify the result against held-out data and physical sanity checks before it is allowed to count.
- 04On a missed target, trace the root cause with Agentic RAG over prior runs, datasheets and literature, then a judge weighs the trade-off and decides advance, re-optimize or reject.
- 05Write the verified outcome, rationale and metrics back to the design record so the next iteration starts from ground truth.
05Tooling
Simulation, solvers & frameworks.
Get started