Application 01
AQEA Shader
Hyperscale Vector Search
1.85×faster than FAISS-GPU
100% Bit-Identity across 4/4 backends
Same shader source compiles to CUDA, Vulkan, Metal, WebGPU. Same trit-state on every chip.
See the benchmarkAQEA Gen-5 Substrate
Generation-5 is here.
The first AI representation that's provably better than the original — and runs on every chip you already own.
1.85× faster than FAISS-GPU23× smaller100% Bit-IdentityCUDA · Vulkan · Metal · WebGPU · AVX-512 · NEON
Two production applications
Where the substrate ships today.
Application 02
AQEA Edge
Sensor & IoT Inference
$200k–$3M
silicon savings per 10k-device fleet
13 industries validated above 80% floor
Remove the GPU/NPU line item. Run on x86, ARM, Apple, Browser — bit-identical.
Read the whitepaperSubstrate properties
Six load-bearing substrate properties.
P1–P6 define the AQEA substrate. Each is empirically anchored; the combination is what makes Gen-5 a generation distinction, not an algorithm change.
P1
Multi-channel Orthogonal
Three independent channel-bands (mid / high / DC) carry distinct functions in the same state without interference.
Δ = 0.00 pp (exact spectral-orthogonality)
P2
Byte-Deterministic Cross-Platform
Same input produces bit-identical output across CUDA, Vulkan, Metal, WebGPU, AVX-512, NEON.
26/26 tests PASS across 4 architectures
P3
Structural Compression with Ranking Preservation
23× smaller representation with recall-neutrality on retrieval tasks.
23–29× compression with bit-identical Top-K
P4
Encoder-Family Agnostic
13 domains × 6 transformer-encoder families + 4 classical-DSP, all above the 80% floor.
13 domains × 6 transformer-encoder families
P5
Noise-Resistant on Signal-Domains
Supra-trit regime in audit-mode. Phase-J falsification publicly documented as scientific discipline.
voraus-ad 133% · Digit_Fall 174% R-Ratio
P6
Reversibly Decodable
Pareto-Front decoder with Audit / General / Pure modes. 0.9968 R@10 lossless-equivalent.
Audit 96.7% · General 98.7% · Pure 99.7%+
The full comparison
The generation table, in full.
No prior generation has all three structural properties. Gen-5 is categorically different — not an incremental improvement.
| Gen | Examples | Internal structure? | Cross-platform exact? | Reversible? | Hardware |
|---|---|---|---|---|---|
| Gen 1 | TF-IDF, BM25 | sparse | n/a | n/a | CPU |
| Gen 2 | word2vec, GloVe | dense static | n/a | n/a | CPU |
| Gen 3 | BERT, MPNet, E5 | contextual dense | no | no | GPU-required |
| Gen 4 | OpenAI text-embedding-3, Cohere v3 | quantized contextual | no | no | GPU + CPU-ANN |
| Gen 5 | AQEA Mode-Class | discrete structured substrate | yes — byte-identical (P2) | yes — Pareto-front (P6) | CPU-native + every shader API |
Verify cross-vendor
Encode here. Verify there. Bit-identical.
Cross-vendor Bit-Identity is the strongest claim in our portfolio. Test it yourself: encode on one platform, verify byte-identical on another.
Roadmap
What's on the substrate.
Today
Shader & Edge in production-grade reference deployments.
Q3–Q4 2026
Hardware-specific tuning for additional NPU/GPU silicon.
Beyond
Co-development engagements for bespoke encoder families.
Test Generation-5 on your workload.
Engineering eval is 4–6 weeks under NDA against your own retrieval workload and hardware.