Projects

The Architect

Production (PyPI)
Public • Open-Source
  • Autonomous development framework designed for agentic AI programming tools to manage planning, execution, iterative retrying, code review, and persistent memory without human intervention.
  • Maintains a persistent project state via ARCHITECT.md to capture cumulative decisions, constraints, and lessons across sessions.
  • Employs multi-faceted completion detection and integrated circuit breakers to prevent agent deadlock and hallucinated outputs.
  • Supports headless operation and continuous integration workflows with a live tmux dashboard for real-time monitoring.
  • Published on PyPI, installable via pip, and compatible with LLMs such as Claude Code and OpenCode.

MHIL

Concept & Early Prototyping (TRL-5)
Semi-Stealth • Non‑Profit
  • Machine Human Intelligence Lab (MHIL) is a non-profit research institute developing benchmark suites and evaluation methodologies targeting true machine intelligence beyond narrow LLM task accuracy.
  • Constructs human-aligned test batteries assessing agent-level capabilities including memory, planning, robustness, and autonomous problem-solving under uncertainty.
  • Produces open evaluation matrices for comparative analysis of complex AI systems across models, vendors, and architectures.
  • Orients development towards reproducible scorecards, capability tiers, and safety-critical performance thresholds suitable for pre-superintelligent systems and real-world deployment standards.

MazeByte

SOTA Proof of Concept (PoC)
Semi-Stealth
  • A research project demonstrating a fully autonomous AI agent capable of end-to-end ETL data pipeline automation without human oversight.
  • Implements adaptive learning via infant-inspired feedback loops employing reward and pain mechanisms, supplemented with resource-seeking and self-preservation behaviours.
  • Generates modular pseudo-code components responsible for long-term memory, search functions, optimisation, and orchestration of data workflows.
  • Built around the proprietary MBI model, MazeByte aims to supplant traditional manual data engineering by delivering persistent self-maintained automation solutions.

SightBit

SaaS (TRL-9)
Public
  • Production-grade computer vision SaaS platform delivering real-time detection of drowning risks, human and vessel presence, and flood alerts leveraging standard CCTV footage alone.
  • Eliminates reliance on specialised sensors, edge computing hardware, or bespoke per-site calibration.
  • Utilises an advanced panoptic segmentation model enhanced with proprietary data augmentation pipelines to substantially improve recall rates for small objects by 18 percentage points.

AdaptiveBridge

Production-ready (TRL-8)
Public • Open-Source
  • Open-source Python library designed to enhance machine learning model resilience by dynamically predicting and imputing missing features during inference in production.
  • Supports configurable thresholds for feature importance, correlation analysis, and accuracy criteria, integrating custom logic and automated distribution adjustments.
  • Provides a drop-in estimator interface with fit-transform methodology for seamless embedding into existing ML workflows.
  • Manages complex feature dependencies, including mandatory, deviation, and leveled features, to ensure robust model performance.
  • Delivered with comprehensive documentation, unit tests, and continuous integration, published under a permissive open-source license on PyPI.

Jhive

Production (TRL-7)
Public • AIforGood
  • Django-based application utilising the OpenAI API for real-time detection and classification of antisemitic content.
  • Features a detailed reporting interface to facilitate content analysis and stakeholder review.
  • Offers RESTful API endpoints with built-in rate limiting to support scalable integration into external systems.
  • Incorporates rigorous input validation and sanitisation mechanisms to mitigate injection attacks and ensure data integrity.