ai engineer

codengo

// building intelligence from scratch

principal software engineer crafting neural architectures in pure go. no frameworks. no shortcuts. just mathematics, algorithms, and four decades of obsession with code.

40+

years coding

11+

years in go

∞

ai obsession

01 — focus areas

what i build

[ neural architectures ]

deep learning from scratch

implementing mlps, kans, esns, and transformers in pure go. backpropagation, gradient descent, attention mechanisms... all hand-coded.

mlp kan esn transformers

[ inference systems ]

llm inference tooling

building llama.cpp wrappers, optimized inference pipelines, and custom serving solutions. production-grade ai with go's performance.

llama.cpp gguf quantization

[ vector systems ]

embeddings & retrieval

custom vector databases, semantic search engines, and rag pipelines. turning unstructured data into intelligence.

embeddings hnsw rag

[ activation research ]

experimental architectures

exploring novel activation functions, learnable splines in kans, reservoir computing dynamics. pushing the boundaries.

b-splines gelu swish

[ high-performance go ]

systems engineering

concurrent processing, memory optimization, simd operations. making go fast enough for neural network training.

goroutines simd optimization

[ model architecture ]

experimental models

designing and testing novel neural network architectures. combining classical approaches with modern insights.

research architecture experimentation

02 — implementations

neural projects

nn—001

multi-layer perceptron

go • pure math • backpropagation

feedforward neural network with configurable layers, multiple activation functions, and optimized matrix operations. training via gradient descent.

nn—002

kolmogorov-arnold network

go • b-splines • learnable activations

novel architecture with learnable activation functions on edges. implementing the kolmogorov-arnold representation theorem in code.

nn—003

echo state network

go • reservoir computing • temporal patterns

reservoir computing implementation with sparse random connectivity. efficient training through linear regression on readout weights.

nn—004

llama.go

go • llama.cpp • cgo • inference

high-performance llm inference wrapper with gguf support, context management, and optimized memory handling for production deployments.

nn—005

vectordb

go • mmap • unsafe • zero-allocation

high-performance vector database utilizing mmap and unsafe for zero-allocation operations. hnsw indexing with memory-mapped persistence for production-scale similarity search.

transformer.go

1// attention is all
                            you need — in go
2func  (t
                        *Transformer) Attention(q, k, v
                        [][]float64) [][]float64  {
                    
3    scores  := t.MatMul(q, t.Transpose(k))
4    scaled  := t.Scale(scores, 1.0/math.Sqrt(t.dim))
5    weights := t.Softmax(scaled)
6
7    return 
                        t.MatMul(weights, v)
8}

03 — evolution

the path to ai

1981

the genesis

father builds a zenith heathkit h89. the first computer enters our home. a spark ignites in a child's mind.

1984

first lines of code

age seven. msbasic | qbasic. teaching myself in the green glow of a crt monitor. the obsession begins.

1990s

deep systems

c. c++. diving into memory management, pointers, the metal beneath the abstractions.

2013

go enters

discovered golang. simplicity meeting power. concurrency as a first-class citizen. home.

2016 — 2024

enterprise scale

building high-performance apis at top fortune 50 companies. millions of requests. zero excuses.

2024

the ai pivot

building mlp, kan, esn neural networks in pure go. no frameworks. just mathematics and determination.

now

convergence

transformers. llm inference. vector databases. dedicated to pushing the boundaries of ai in go.

codengo

what i build

neural projects

the path to ai

let's build intelligence