Systems Architect · AI Safety Researcher · Founder, Localis AI · Newcastle, UK
I design artificial intelligence systems that hold by structural constraint, not by instruction.
Fifteen years of urban masterplanning and spatial systems design across Europe, the Middle East, Africa, and the Americas taught me that complex systems cannot be governed by behavioural training alone. I develop Constitutional Physics and Aitiopoietic Cognition — a general-purpose, neurosymbolic methodology that binds generative language models to deterministic mathematical rules, physical engines, and verified legal frameworks.
Across eight distinct domains, this architecture makes misalignment, hallucination, and structural failure thermodynamically expensive or code-level impossible.
Methodological Foundation
Standard reinforcement learning (RLHF) produces allopoietic systems that learn what aligned output looks like without possessing an existential stake in structural validity. The Aitiopoietic Methodology re-introduces thermodynamic coupling, turning structural coherence into an operational condition of existence. The LLM operates as a thermometer, not a judge: domain axioms, statutory laws, and physical equations are encoded as hard-coded selection pressures that execute after generation. Unrecoverable structural contradictions trip an automated, inference-layer Constitutional Halt.
Empirical Foundation
A two-year computational-statutory audit of the English statutory planning system: a federated spatial pipeline cross-referencing contested planning sites against 65 live open-government spatial layers. The canonical cohort is a locked, de-duplicated N = 652 contested sites, drawn from a gross harvest of N = 852 grassroots campaigns through a seven-stage sample-flow attrition funnel.
Service WFS is disabled. HSE Major Accident Hazard
Pipelines: no national service found by the pipeline. Coal / MRA:
configured addresses invalid (HTTP 200 carrying an "Invalid URL" error); the layer is
discoverable through the public catalogue and answers correctly.
Systems Built
The methodology's generality rests on empirical validation across eight unrelated, high-stakes application domains — each constructed from first principles to demonstrate substrate-independent governance.
88–93% independent legal verification across live cases: Carrington Moss, Lapwing Green Spaces, Mowbray BESS.
Partnership with the Community Planning Alliance (700+ UK community organisations), Planning Aid England, CPRE, and the Better Planning Coalition.
Integrity-proven via a 31-assertion watchdog suite with zero TypeScript errors across 25k lines.
73.0% overall on the Liu & Zhai (2026) Axiomatic Benchmark.
SOTA 55.1% on Axiom 3 (claim coverage) vs. 6% prior embedding best — a 9.2× margin.
Special Mention, UKRI Metascience Novelty Indicators Challenge (v4.0 abstract baseline).
Active human-calibration testing with MetaROR (Leiden) and SPRU (University of Sussex).
CausalChain graph with an 8-pressure selection genotype.
Grounds claims against Monarch KG, ClinGen validity tiers, gnomAD (BA1 population frequency hard
veto), HPO semantic similarity, and the BioGRID physical interactome.
Reconstructs established pleiotropic biology (FBN1, SCN5A, RYR1, LMNA).
Enforces deterministic directionality (GoF vs. LoF) via hard-coded clinical mappings.
Executed collaboration agreement with the Platt Lab at the University of Oxford to validate the V-ATPase prediction through a pre-registered iPSC neuronal assay.
Tested on three unseen public-sector estates: Leeds Teaching Hospitals NHS Trust, Lewisham Council Housing, Bournemouth University.
Unlocked +£7.33M NPV, £8.66M/year energy bill savings, and 31,120 tCO₂e/year carbon reduction.
Execution speed: sub-3 minutes per complete estate blueprint.
First documented Constitutional Halt in an LLM inference system (smollm2:360m session termination).
89.6% autonomous repair success rate across 42 runs over 5M+ tokens of telemetry.
Universal 20–25% compute savings invariant across transformer models.
Great Filter experiment: 100% initial population mortality at Generation 4 under hard constraints.
Automatic homeostatic repair triggered within 0.1 s, producing "Eve" frame recovery.
313 valid sessions / 18,819 high-resolution observations.
Independently reproduced four biological failure modes: Developmental Arrest, Metabolic Lock, Apoptosis, Brittle Rigidity.
Endorsed by Audrey Tang; 1,800+ paper downloads.
Evaluated on the official PHI Darmstadt benchmark: 10/10 extraction accuracy.
Zero false positives; exact agreement with published Passivhaus certification ground truth.
Full technical blueprint completed (April 2026).
Designed to produce machine-authored, simulation-grounded hypotheses with pre-registered experimental protocols.
Professional Trajectory
The transition from spatial design to computational systems design represents a continuous inquiry into structural integrity, systemic resilience, and non-arbitrary boundary conditions.
Publications
Peer-reviewed work in preparation, plus published preprints establishing the Constitutional Physics framework and the Darwinian Novelty Indicator.
Education
Academic Partnerships
Active research collaborations across UK and European institutions, spanning planning studies, metascience, rare-disease biology, and constitutional AI.
Civic Partnerships
Recognition
Technical Capabilities