Transformer encoders
Six transformer families validated — BGE, WavLM, ESM-2, BiomedCLIP, codet5p, CLAP — across text, speech, protein, medical imaging, code and music.
AQEA Gen-5 · Edge
Remove the $200k–$3M silicon line item.
Sensor-stream inference without dedicated AI accelerators — same accuracy, full audit trail, cross-vendor deterministic.
13 domains validatedNo GPU requiredARM Cortex-M4 capableReversible decoding
Section 01
The BOM-cost story.
A 10,000-device fleet pays the silicon line item once per device, forever. Removing it changes unit economics, not just margin.
| Silicon | Per Unit | 10k-device Fleet |
|---|---|---|
| Jetson Orin Nano | $300 | $3,000,000 |
| Coral TPU | $60 | $600,000 |
| Hailo-8 | $80 | $800,000 |
| AQEA Edge | $0 | $0 |
AQEA Edge runs on the CPU and on-board NPU your device already ships with.
Section 02
13-domain validation.
Empirical results across 13 domain × encoder combinations, six transformer-encoder families plus four classical signal-processing pipelines.
| Domain | Encoder Family | AQEA R@10 Ratio |
|---|---|---|
| Text (msmarco) | BGE-large | 97.4% |
| Speech | WavLM | ≥80% |
| Protein | ESM-2 | ≥80% |
| Medical Imaging | BiomedCLIP | ≥80% |
| Code | codet5p | ≥80% |
| Music | CLAP | ≥80% |
| Spectral FFT | classical-DSP | ≥80% |
| Image DCT | classical-DSP | ≥80% |
| Multispectral Bands | classical-DSP | ≥80% |
| Mass-Spectrometry | classical-DSP | 96.7% (Phase J PASS) |
| Robotic-Arm Anomaly (voraus-ad) | classical-DSP | 133% (EXCEEDS Float) |
| Wearable Fall (Digit_Fall) | classical-DSP | 174% (EXCEEDS Float) |
| Industrial-Sensor #3 | classical-DSP | EXCEEDS Float |
Section 03
Encoder-family-agnostic.
Classical DSP pipelines
Four classical signal-processing pipelines — FFT-spectral, DCT, multispectral, mass-spec — through the same encoder interface. Same compression, same Top-K guarantees.
Section 04
Cross-vendor bit-deterministic.
Hardware validated: x86 AVX-512 (Sapphire Rapids, AMD EPYC), ARM NEON (Apple M3 Pro/Max), and four GPU backends (Vulkan / Metal / WebGPU / CUDA). Same shader source. Byte-identical results.
x86 AVX-512
ARM NEON
NVIDIA CUDA
Apple Metal
WebGPU
Section 05
Audit-trail reversible decoding.
For medical devices, autonomous systems, financial trading and any regulated-industry edge deployment, you need to prove what your inference was based on.
Mode 01
Audit
96.7%
Retrieval-equivalent, ≥0.90 per-vector cosine reconstruction-fidelity.
Mode 02
General
98.7%
Retrieval-equivalent, balanced fidelity and discrimination.
Mode 03
Pure Retrieval
99.7%+
Retrieval-equivalent, tighter discrimination, leanest decoder.
Decoder is a deployment-dial — select per-workload, no re-encoding of the substrate required.
Section 06
Engineering engagement.
Step 01
Engineering Eval
4–6 weeks
On your sensor stream.
Step 02
Integration Pilot
In firmware
Shadow-mode on your device.
Step 03
Co-Development
Hardware tuning
Bespoke encoder families.
Move sensor inference off dedicated silicon.
We respond to engineering and co-development inquiries within one business day.