Client work · Regulatory model
A compliance SLM,private to the client’s cloud.
Overview
It reads one regulatory article and returns the risks and controls that bind, or nothing when the text does not. A reviewer still signs. Built for a client in a regulated industry, it runs inside their own AWS account.
27B parameters, one square per billion
94 in 100 answers match the reviewer
The model
A 27B specialist model, developed from an expert-labelled workflow and built with supervised fine-tuning and reinforcement learning. Synthetic data covers the rare labels. It is scored on a held-out gold set, with the expert labels withheld from training.
- Expert-labelled work
- Synthetic data for rare labels
- Fine-tuned, then reinforced
- Held-out gold set
What the model does
- Returns the risks a regulatory article creates
- 94% match
- Maps every risk to the control that answers it
- 93% match
- Emits output the downstream system can parse
- 100%
- Covers rare labels through synthetic data
- Synthetic
- Returns nothing when the text does not bind
- Guardrail
- Leaves the client’s own AWS account
- Never
- Signs the compliance opinion
- Never
Stack
Model
- 27B specialist model
- Supervised fine-tuning
- Reinforcement learning
- Synthetic data
- Held-out gold set
AWS
- Control Tower
- Organisations
- IAM
- Amazon S3
- SageMaker
- Amazon VPC
- CloudWatch
- Budgets
Hello from Malmö
Adamass
Come with the work that feels heavy. We will make it light.
- IKEA
- Nasdaq
- Tele2
- Thomson Reuters
- The Guardian
- Liverpool FC
- Stabenfeldt
- Eyeradar

