Projects

The Architect

Production (PyPI)
Public • Open-Source
  • Autonomous lifecycle management for agentic AI coding tools—plans, executes, retries, reviews, and retains knowledge with zero supervision.
  • Implements persistent project memory via ARCHITECT.md to accumulate decisions, constraints, and insights across sessions.
  • Multi-signal completion detection and robust circuit breakers prevent agent deadlocks and hallucinations.
  • Supports headless and CI workflows with live tmux dashboard for flexible deployment.
  • Released on PyPI with compatibility for Claude Code and OpenCode, enabling seamless installation via pip.

MHIL

Concept & Early Prototyping (TRL-5)
Semi-Stealth • Non‑Profit
  • Machine Human Intelligence Lab (MHIL) is a non-profit research institute developing benchmarks and evaluation protocols that transcend narrow LLM tasks to assess true machine intelligence.
  • Designs human-aligned test batteries for agents, tool-using systems, and composite architectures, evaluating capabilities like memory, planning, robustness, and autonomous reasoning under uncertainty.
  • Creates an open evaluation matrix facilitating direct comparison across models, vendors, and architectures, targeting future ‘pre-superintelligent’ systems.
  • Focuses on reproducible scorecards, defined capability tiers, and safety-critical thresholds aimed at real-world deployment standards.

MazeByte

SOTA Proof of Concept (PoC)
Semi-Stealth
  • Proof-of-concept demonstrating fully autonomous, self-maintained AI agents for end-to-end ETL automation in data engineering.
  • Employs infant-like feedback mechanisms based on reward and pain signals, combined with resource-seeking and self-preservation behaviours, enabling continuous adaptive learning.
  • Generates pseudo-code modules supporting memory, search, optimisation, and pipeline orchestration.
  • Built around a proprietary MBI internal model, MazeByte aims to eliminate manual intervention in data pipeline creation and maintenance by treating data systems as adaptive, agent-driven entities.

SightBit

SaaS (TRL-9)
Public
  • Real-time computer vision platform for drowning prediction, human and vessel detection, and flood-risk alerting, operating on standard CCTV feeds without requiring specialised sensors or site-specific calibration.
  • Utilises a panoptic segmentation model enhanced by a proprietary data augmentation pipeline, significantly improving recall rates for small-object detection.
  • Engineered for operational reliability and scalability in safety-critical aquatic environments, delivering continuous monitoring with immediate risk detection capabilities.

AdaptiveBridge

Production-ready (TRL-8)
Public • Open-Source
  • Open-source Python library ensuring robust ML model predictions despite missing features at inference time.
  • Predicts and reconstructs missing data dynamically, maintaining model performance in real-world, imperfect environments.
  • Features configurable thresholds for feature importance, correlations, and accuracy metrics, supporting custom logic and automatic distribution handling.
  • Provides a drop-in estimator interface integrating seamlessly with existing ML pipelines via fit and transform methods.
  • Manages complex feature dependencies including mandatory, deviation, and hierarchical features to optimise model resilience.

Jhive

Production (TRL-7)
Public • AIforGood
  • Production-grade Django web application leveraging the OpenAI API for real-time antisemitic content detection and analysis.
  • Offers a detailed reporting interface for content review and moderation workflows.
  • Provides secure RESTful API endpoints featuring rate limiting and thorough input sanitisation to prevent injection attacks and misuse.
  • Engineered for reliability and safety in sensitive content moderation contexts, designed specifically to address antisemitic content rather than general toxicity.