We compress LLM KV cache. Measured: 8× at 3-bit with 0.43% quality delta (GPU-validated); 22× live end-to-end on a 5-tenant workload, GPU-verified this week (quantization × prefix dedup, decomposition always stated); 16.89× tenant density on a single V100-32GB. Validated across Mistral/Qwen/Llama with zero re-tuning. 90+ patents pending; design-partner pipeline; seed round in motion.
The gap: our throughput path is a reference implementation. First job is the fused Triton/CUDA dequant+attention kernel — turning a validated 22× memory result into production speed. Then you own the serving stack.
Process: call → NDA → paid 1–2 week trial on the real kernel → decide together. We report negative results straight; you should want to work somewhere that does.