Y Combinator S26 Applicant

2.15x Faster LLM Inference
Zero Quality Loss

CDLaC accelerates context processing for enterprise LLM deployments, reducing compute costs while improving benchmark scores.

2.15x
Prefill Speedup
16,904 tok/s vs baseline
1.42x
Decode Speedup
37.5 tok/s vs baseline
+10.5
Quality Gain
ARC-Easy benchmark
View Benchmarks Request Evaluation

The Hidden Cost of Reading

Enterprise LLM deployments spend most of their GPU compute on prefill — processing context before generating a single output token.

60-80%
of GPU time spent on prefill, not generation
35s+
to process a single 128K-token document
O(L²)
attention cost scales quadratically

Fast Read, Standard Write

CDLaC compresses context during ingestion, reducing attention cost by 4x. When it's time to generate, we restore full resolution for quality output.

Standard: [L tokens] → [L² attention] → [output]
CDLaC:    [L tokens] → [L/2 compressed] → [(L/2)² attention] → [restore] → [output]
Drop-in Integration
Works with existing transformer models
Complementary
Stacks with vLLM, PagedAttention, FP8
Quality Preserved
Actually improves benchmark scores

Verified Performance

All measurements on NVIDIA A100-80GB, January 2026

Speed Benchmarks

Context CDLaC Baseline Speedup
8K tokens 16,904 tok/s 7,866 tok/s 2.15x
32K tokens 11,576 tok/s 5,821 tok/s 1.99x
64K tokens 8,074 tok/s 4,102 tok/s 1.97x
128K tokens 5,944 tok/s 2,654 tok/s 2.24x

Quality Benchmarks

Benchmark CDLaC Baseline Delta
LAMBADA 70.97% 65.57% +5.40
ARC-Easy 80.47% 69.95% +10.52
PIQA 79.38% 76.77% +2.61
Winogrande 73.80% 70.48% +3.32

Full methodology and reproducible scripts on GitHub →

Where CDLaC Shines

Acceleration benefits scale with context length — exactly where providers need help most

📄

Document Analysis

Process lengthy contracts, reports, and research papers 2x faster

🗂

RAG Pipelines

Ingest retrieval context at scale without proportional cost increase

💻

Code Understanding

Analyze entire repositories for review, refactoring, documentation

⚡

Real-time Apps

Hit latency SLAs on long-context requests without over-provisioning

Built by an Optimization Engineer

MH

Mike Holford

Founder & CEO

20+ years turning computational bottlenecks into competitive advantages. CDLaC applies decades of efficiency engineering to the $50B+ LLM inference market.

13
US Patents
20+
Years Experience
🔗 LinkedIn 💻 GitHub

Ready to Accelerate?

Get benchmark access for your specific workloads

Request Evaluation

Investor Portal

Access detailed materials with your investor code

Invalid access code

Email
mike@codynamicslab.com
Location
Gilbert, AZ
Entity
Delaware C-Corp