Building Blocks

Cohesive.AI

Cohesive.AI is the AI and numerical-computation block for inference, training, dataset materialization, model artifacts, vector search, text processing, and numerical primitives.

Core Idea

The long-term direction is an AI IR: a portable semantic representation of model structure, tensor operations, training plans, inference plans, feature extraction, and evaluation logic that can be compiled or projected onto concrete runtimes.

Why It Exists

AI systems often mix model-specific inference APIs, tokenizers, feature extraction, vector indexes, training data projection, training code packaging, job submission, model storage, evaluation, promotion, and workflow orchestration in one runtime.

Cohesive.AI separates those concerns into semantic contracts so each part can be trained, registered, queried, orchestrated, and projected without every component depending on one vendor stack.

Inference

Inference is modeled by capability rather than a single generic Predict method. Models can expose embeddings, classification, extraction, generation, reranking, scoring, relation inference, or domain-specific numerical operations through typed semantic interfaces.

Training and Artifacts

Training plans can describe datasets, feature extraction, model configuration, evaluation metrics, runtime requirements, promotion criteria, and generated artifacts.

Model artifacts can be registered with lineage: which data produced them, which training plan ran, which metrics were observed, and which inference surfaces the artifact supports.

Vectors and Text

AI integrations often need embedding models, vector indexes, semantic ontologies, tokenization, chunking, normalization, text classification, and search-time reranking.

Cohesive.AI keeps these as composable semantic pieces that can connect to relations, storage, presentation, and processes.

Backend Targets

PyTorch / JAX

Role: Training and model authoring.
Interpretation: Lower AI plans into Python-oriented training runtimes.

ONNX Runtime

Role: Portable inference.
Interpretation: Run registered artifacts through stable inference contracts.

ILGPU / local CPU

Role: Numerical execution.
Interpretation: Evaluate kernels or small models in local .NET-oriented runtimes.

Cloud ML systems

Role: Managed training and serving.
Interpretation: Submit jobs, register artifacts, and capture lineage through adapters.

Vector stores

Role: Semantic retrieval.
Interpretation: Bind embeddings and retrieval plans to backend-specific indexes.

Target	Role	Interpretation
PyTorch / JAX	Training and model authoring.	Lower AI plans into Python-oriented training runtimes.
ONNX Runtime	Portable inference.	Run registered artifacts through stable inference contracts.
ILGPU / local CPU	Numerical execution.	Evaluate kernels or small models in local .NET-oriented runtimes.
Cloud ML systems	Managed training and serving.	Submit jobs, register artifacts, and capture lineage through adapters.
Vector stores	Semantic retrieval.	Bind embeddings and retrieval plans to backend-specific indexes.