Artificial Intelligence · Sovereign AI · Data and Knowledge

ADG LLM

Language models engineered for your domain — and your infrastructure.

In active developmentADG LLM is an active model-engineering initiative. Model availability is confirmed per engagement.

Overview

Model engineering, fine-tuning and private inference.

ADG LLM is ADGTech's model-engineering initiative: selecting, adapting, evaluating and serving language models that run under your control. Today, the work centres on fine-tuning open-weight foundation models for specific domains and serving them through private inference — not training foundation models from scratch.

Interactive preview

Explore ADG LLM

An interactive demonstration using sample data. It illustrates the experience and does not connect to live systems.

ADG LLM architecture

Reference stack · select a layer

Illustrative architecture

Evaluation

Applies across all layers

Task benchmarks
Baseline comparison
Regression checks

Responsible-AI controls

Applies across all layers

Content filters
PII handling
Human review

- Consumers call models through one internal API rather than integrating with each model separately
- OpenAI-compatible endpoints can ease migration of existing applications
- Per-application keys, quotas and usage reporting
- Routes each request to an approved model based on workspace, task and data classification
- Blocks routes to externally hosted models unless explicitly approved
- Request logging for audit, with configurable retention
- GPU inference servers with batching and streaming responses
- Quantisation options to fit models to available hardware
- Runs on-premises or in a private cloud environment, as agreed per deployment
- Records which base model and adapter version is approved for which use
- Evaluation results and approvals are attached to each version
- Supports rollback to a previously approved version
- LoRA / QLoRA adapters trained by ADGTech on curated, approved datasets
- Adapters are small and swappable — the base model is not retrained
- Each adapter is evaluated against the unmodified base model before approval
- Chosen against your task, languages, licence terms and hardware
- Used under the publisher's licence; ADGTech does not train foundation models from scratch
- Can be served unmodified or combined with fine-tuned adapters

Model options compared

ADGTech adapts existing open-weight foundation models through fine-tuning. It does not train foundation models from scratch.

Attribute	ADG fine-tunedOpen-weight models adapted by ADGTech on your data	Open-weightServed privately, unmodified	Externally hostedThird-party APIs, only where approved
Where it runs	Your infrastructure or a private environment agreed with ADGTech	Your infrastructure or a private environment agreed with ADGTech	The provider's infrastructure
Who controls weights	Base weights from the publisher under its licence; adapter weights trained for you, with ownership agreed per engagement	Publisher releases the weights; you host and run an unmodified copy under its licence	The provider — weights are not accessible
Data boundary	Training data and prompts stay within your boundary	Prompts and responses stay within your boundary	Prompts leave your boundary and are handled under the provider's terms
Customisation	High — adapters tuned to your terminology, formats and tasks, then evaluated	Prompting, retrieval (RAG) and tool use; no weight changes	Prompting and retrieval; any tuning depends on the provider's options
Typical use	Domain tasks where a general model underperforms on evaluation	General assistants, RAG answers and a baseline for comparison	Approved, lower-sensitivity tasks or capability comparison

ADG fine-tuned

Open-weight models adapted by ADGTech on your data

Where it runs: Your infrastructure or a private environment agreed with ADGTech
Who controls weights: Base weights from the publisher under its licence; adapter weights trained for you, with ownership agreed per engagement
Data boundary: Training data and prompts stay within your boundary
Customisation: High — adapters tuned to your terminology, formats and tasks, then evaluated
Typical use: Domain tasks where a general model underperforms on evaluation

Key features

Capabilities

Model engineering

Base-model selection against your task, language and hardware
Domain-specialised fine-tuning (LoRA / QLoRA)
Prompt, tool-use and structured-output tuning
Quantisation for efficient private inference

Evaluation & lifecycle

Task-specific evaluation datasets
Benchmarking against baseline models
Model registry and versioning
Approval gates before promotion to production

Serving & control

Private inference endpoints
Enterprise APIs with authentication
Model routing between private and approved external models
Responsible-AI controls, logging and review

Workflow

How it flows

01Select base model
02Prepare dataset
03Fine-tune
04Evaluate
05Approve
06Serve
07Monitor

Use cases

Where it delivers value

Domain language models
Adapt an open-weight model to your terminology, document formats and tasks.
Private inference
Serve models inside infrastructure you choose so prompts and outputs stay within your boundary.
Model evaluation
Measure whether a fine-tuned model actually outperforms a baseline on your tasks before rollout.

Architecture overview

How it's built

Models
Open-weight base models, ADG fine-tuned adapters, and optionally approved hosted models.
Training
Fine-tuning pipelines with experiment tracking and reproducible configs.
Registry
Versioned model artefacts with evaluation results and approvals.
Serving
Private inference runtime behind an authenticated API gateway.

Integrations

Enterprise RAG
ADG Chat
Agent orchestration
OpenAI-compatible APIs
Customer applications

Deployment options

On-premise GPU servers
Private cloud
Dedicated hosted infrastructure
Hybrid with approved external models

Available deployment options are confirmed per engagement.

Works well with

See ADG LLM in action

Book a guided demo with the ADGTech team, tailored to your workflows and environment.

Request a demo Contact sales