A model that knows your business
Public LLMs know everything about the world and nothing about your organisation. Model Factory generates fine-tuned small language models (SLMs) — trained on your knowledge graph, documents and approved agent decisions, specialised for one workflow, benchmarked against the base model, and served privately. World-first: a model built from your own enterprise knowledge, not the public internet.
Fine-tuned on your knowledge. Tuned to your workflow.
A Model Factory model is a behavioural specialist — it learns how your organisation runs a specific process. Facts still come from your Knowledge Graph and RAG at run time; the model learns the reasoning, format and tone your teams have already approved.
-
Trained on your knowledge
Your knowledge graph, ontologies, documents and approved agent traces — not the public internet.
-
Specialised per workflow
One tuned model per workflow or agent — a right-sized SLM for the narrow, high-volume tasks.
-
Private & owned
Training data never leaves your environment; the model runs on your own infrastructure.
-
Improves over time
Retrain on newer approved runs — the model gets measurably better the more the workflow is used.
From a workflow to a ready-to-run model
Pick a workflow, choose an open model and method, and Model Factory emits ready-to-run training, benchmarking and serving notebooks — with a governance model card.
1 · Select & configure
Choose a workflow, a base model and a method (LoRA / QLoRA). Optionally add a human-preference (RL) job.
2 · Prepare data
Build a training corpus from traces, documents, and your knowledge graph — locally, so data stays with you.
3 · Train & benchmark
Fine-tune on Colab or your GPU, then benchmark the tuned model against the base on a held-out test split.
4 · Serve & route
Host via Ollama or vLLM behind an OpenAI-compatible endpoint; route an agent to it with no code changes.
Any open model. Any method.
Choose from leading open-source model families and the precision that fits your GPU — from full-precision LoRA to 4-bit QLoRA that trains on a free Colab GPU.
Qwen 2.5
Llama 3.x
Mistral
DeepSeek
Gemma 2
Phi-4
LoRA · 16/32-bit
QLoRA · 4/8-bit
Five ways to build the training data
Mix freely — every source becomes the same instruction-tuning format, prepared locally so nothing leaves your environment.
Workflow & agent traces
Approved runs and human overrides from your deployed workflows.
Documents
PDF, Word, Excel, PowerPoint and more, turned into grounded Q&A.
Knowledge Graph & Ontology
Your entities, relationships and 700+ ontologies as Q&A pairs.
Vector store Q&A
Existing embeddings and RAG chunks converted into training pairs.
Live KG / RAG Q&A
Ask your own Knowledge Graph hundreds of questions via API — Alphient answers them.
Human preference (DPO)
Reviewer corrections become an optional RL job that aligns the model to your judgement.
Own the model, own the moat
Lower cost & latency
Run high-volume workflow tasks on a small private model instead of a frontier API.
Data sovereignty
Train and serve entirely on your own infrastructure — ideal for regulated industries.
No lock-in
Open models, OpenAI-compatible serving, per-agent routing — mix tuned and frontier models.
Governed & benchmarked
Every model ships a scorecard and model card — promote only if it beats the base.
See Model Factory in action
Turn your workflows and knowledge into a private, fine-tuned model — trained, benchmarked and served from your own data.