Systems & Reference Architectures
Practical architecture, control, evaluation, and assurance systems in code.
Every repository pairs executable Python systems with architectural decision records (docs/WHY.md / ADRs), automated offline CI, and explicit boundaries.
Core AI systems
Four connected architectures answering one leadership question: how do you scale AI value without losing operational control?
-
Architect & price
ai-systems-architecture
Which deployment boundary meets sovereign rules, and what is the true monthly TCO once human checking and rework are priced?
Deterministic architecture validator (10 sovereignty, residency, ZDR, latency, and security gates), tiered model-routing TCO simulator, and 4 executive ADRs.
-
Control execution
agent-guardrails
How do you let an AI agent act without it sending the wrong thing, twice, to the wrong people?
Runtime control layer for AI agents that write or send: allow/draft/approve modes, execution-time re-validation, idempotency, tamper-evident hash-chained audit logs, and a kill switch.
-
Verify quality
evidence-synthesis-eval
Does an AI synthesis say what the sources say, what did it miss, and what did verification cost?
Open evaluation harness built on UK AISI Inspect: faithfulness, importance-weighted coverage, misstatement rates, blinded human-judge agreement (Cohen’s kappa), and checking effort.
-
Govern & assure
ai-assurance-kit
How can a leadership team run proportionate AI assurance in an afternoon rather than a quarter?
Policy-as-code assurance engine turning AI use cases into schema-validated YAML records, 26 explainable checks, executive one-page cards, and NIST AI RMF / UK standards crosswalks.
Clinical AI & data platforms
-
gut-reaction-platform
Reference clinical AI and health-data platform (4 services, synthetic data): clinical NLP phenotyping, fail-closed visual PII audit pipeline, hospital prescribing harmonisation, and genomic ID bridge, with Kubernetes and Terraform definitions.
High-integrity scientific pipelines
Independent reference implementations inspired by biobank-scale clinical genomics workflows. Built from scratch on synthetic data with zero proprietary or participant data.
-
hla-pipeline-manager
Deterministic planning, structural output verification, and SHA-256 checksummed release deployment for multi-batch imputation workflows.
-
genomic-cohort-delivery-pipeline
Governed multi-batch genomic cohort assembly and verified staging pipeline with exclusion tracking, conflict handling, and cryptographic manifests.
Engineering discipline
- Tested Python packages with automated CI, static typing, and
docs/WHY.md/ ADR records. - Explicit operational boundaries: every README documents exact assumptions, threat models, and test scope.
- Synthetic data only — zero real participant data, credentials, or proprietary assets.