Projects
MHIL
Concept & Early Prototyping (TRL-5)Semi-Stealth • Non‑Profit
- Non-profit AI research lab pioneering composite intelligence evaluation beyond narrow LLM benchmarks.
- Crafts human-aligned test batteries assessing agent capabilities including memory, planning, robustness, and autonomous problem-solving in uncertain environments.
- Develops open, reproducible evaluation frameworks enabling direct comparison across models, vendors, and system architectures.
- Establishes future-ready standards with scorecards, capability tiers, and safety-critical thresholds tailored for deployment in pre-superintelligent systems.
MazeByte
SOTA Proof of Concept (PoC)Semi-Stealth
- Advanced autonomous AI agent designed for fully self-maintained ETL and data pipeline automation.
- Utilises infant-like feedback loops modelled on reward and pain to enable adaptive resource-seeking and self-preservation behaviours.
- Constructs pseudo-code modules for memory, search, optimisation, and orchestration to automate end-to-end data engineering operations.
- Employs proprietary MBI model to eliminate manual data maintenance, treating data systems as dynamic adaptive agents.
SightBit
SaaS (TRL-9)Public
- Production-grade computer vision platform delivering real-time drowning risk prediction, human and vessel detection, and flood hazard alerts via standard CCTV feeds.
- Operates robustly without specialised sensors, custom hardware, or site-specific calibrations.
- Implements panoptic segmentation models enhanced by proprietary augmentation techniques to significantly increase recall of small objects critical for safety monitoring.
AdaptiveBridge
Production-ready (TRL-8)Public • Open-Source
- Open-source Python library ensuring machine learning inference resilience by predicting and imputing missing features dynamically.
- Supports configurable thresholds for feature importance, correlation, and accuracy, with extensible custom logic for distribution modelling.
- Offers drop-in estimator compatibility for seamless pipeline integration with fit-transform API.
- Explicitly manages complex feature dependencies including mandatory and conditional inputs to optimise model robustness in production.
- Delivered with comprehensive documentation, testing, and continuous integration under a permissive license.
Jhive
Production (TRL-7)Public • AIforGood
- Django-based AI application leveraging the OpenAI API for real-time detection and classification of antisemitic content.
- Features a secure, scalable architecture with RESTful APIs supporting integration, rate limiting, and robust input sanitisation to prevent injection attacks.
- Provides an interactive reporting interface designed for precise content moderation and analysis
- focusing exclusively on antisemitism detection within operational environments.