BlitzKernels

Production WASM inference kernels for edge AI — built in Rust, compiled to wasm32-wasip2.
Deploy on Cloudflare Workers, wasmCloud, or any WASI runtime. No GPU required at the edge.

Kernel Catalog — 13 Kernels Available

Kernel	Description	Use Case
blitz-embedding	Batched token embedding · mean-pool · L2 norm	RAG, semantic search, feature extraction
blitz-attention	Multi-head attention (MHA) with LSE	Transformer inference block
blitz-kv-cache	Paged KV-cache · LRU/sliding-window eviction	Autoregressive generation
blitz-layernorm-gelu	Fused LayerNorm + GELU activation	BERT/GPT FFN sublayer
blitz-rope	Rotary position embeddings (RoPE)	LLaMA/Mistral position encoding
blitz-fused-mlp	Fused Linear→LayerNorm→GELU→Linear	Full FFN block (GPT-style)
blitz-swiglu	SwiGLU gated activation	LLaMA 2/3, Mistral FFN
blitz-rmsnorm	RMS LayerNorm (no mean subtraction)	LLaMA/Mistral/Gemma normalization
blitz-int8-matmul	INT8 quantized matmul · 4× memory bandwidth	Quantized weight inference
blitz-bf16-matmul	BF16 matmul · H100/TPU native dtype	Mixed-precision LLM inference
blitz-flash-attention	Flash Attention v2 · tiled block · O(N) memory · GQA	Long-context serving
blitz-token-sampler	Greedy / top-k / top-p (nucleus) / temperature sampling	LLM text generation
cc-faculty-wasm	Claude Code cognitive substrate	Agent memory + reasoning integration

→ View full catalog and pricing

---

Why BlitzKernels?

• Pure Rust, zero unsafe — memory-safe by construction, auditable
• WASM-native — runs in Cloudflare Workers, wasmCloud, Fastly Compute, Deno Deploy
• Composable primitives — stack kernels to build full inference pipelines
• No vendor lock-in — wasm32-wasip2 runs anywhere WASI is supported
• Ed25519 signed — every release is cryptographically signed for supply-chain integrity

---

Quick Start

Embedding (semantic search, RAG)

use blitz_embedding::{batch_embed, EmbeddingConfig, EmbeddingTable};

let config = EmbeddingConfig::default();
let table = EmbeddingTable::new(config.vocab_size, config.hidden_dim);
let token_ids = vec![1u32, 42, 100, 7];
let embeddings = batch_embed(&token_ids, &table, &config);
// Vec<Vec<f32>> — L2-normalized, ready for cosine similarity

Multi-Head Attention

use blitz_attention::{fused_mha, AttentionConfig};

let config = AttentionConfig { num_heads: 12, head_dim: 64, ..Default::default() };
let output = fused_mha(&queries, &keys, &values, &config);
// AttentionOutput { data: Vec<f32>, lse: Vec<f32> }

SwiGLU (LLaMA/Mistral FFN)

use blitz_swiglu::{swiglu_forward, SwiGluConfig};

let config = SwiGluConfig { d_model: 4096, d_ff: 11008 };  // LLaMA-7B dims
let output = swiglu_forward(&x, &gate_weight, &up_weight, &down_weight, &config);

RMSNorm (LLaMA/Mistral/Gemma normalization)

use blitz_rmsnorm::rms_norm;

let mut hidden = vec![/* your hidden states */];
let weight = vec![1.0_f32; hidden_size];  // learned scale parameter
rms_norm(&mut hidden, &weight, hidden_size, 1e-6);
// Normalized in-place — faster than LayerNorm (no mean subtraction)

---

Architecture

┌──────────────────────────────────────────────────────┐
│           Your WASI Runtime (CF Workers, etc.)        │
├──────────────────────────────────────────────────────┤
│  blitz-embedding → blitz-attention → blitz-kv-cache  │
│  blitz-rope → blitz-rmsnorm → blitz-swiglu           │
│  blitz-fused-mlp → blitz-layernorm-gelu              │
│  blitz-int8-matmul → blitz-bf16-matmul               │
│  blitz-flash-attention → blitz-token-sampler          │
│  cc-faculty-wasm                                      │
└──────────────────────────────────────────────────────┘
         Pure Rust · No allocator required
         No CUDA · No runtime deps · WASI P2

• Target: wasm32-wasip2 (WASI Preview 2 + Component Model)
• Delivery: Pre-compiled .wasm + Rust source + integration guide
• Signing: Ed25519 (every release cryptographically verified)
• Integration: 30-minute architecture call included with purchase

---

Pricing

Option	Price	Includes
Single kernel	$1,500	Pre-compiled WASM + source + 30-min integration call
Full catalog (13)	$6,500	All 13 kernels + dedicated integration support + priority updates
Support add-on	$200/mo	Priority email, patch releases, architecture review

→ Purchase at blitzkernels.pages.dev
Contact: hello@teamadapt.dev

---

Supply Chain Security

Every kernel release is:

1. Built from auditable Rust source (cargo build --release --target wasm32-wasip2)
2. SHA-256 checksummed
3. Ed25519 signed with our public key (available on request)

Verify a release:

openssl pkeyutl -verify -pubin -inkey blitz-public.pem \
  -rawin -in blitz-embedding.sha256 \
  -sigfile blitz-embedding.blitzpkg.sig

---

License

Source code delivered under commercial license. Contact hello@teamadapt.dev.

---

Built by TeamADAPT — production Rust AI infrastructure.
Running the Nova Collective: 40+ autonomous AI agents, Temporal workflows, 27-tier memory.