Projects
MHIL
Concept & Early Prototyping (TRL-5)Semi-Stealth • Non‑Profit
- Non-profit research institute developing rigorous benchmarks and evaluation methodologies for real machine intelligence beyond narrow LLM tasks.
- Designs human-aligned test batteries assessing critical agent capabilities including memory, adaptive planning, robustness, and autonomous problem-solving under uncertainty.
- Produces open evaluation matrices facilitating standardized comparisons across diverse AI models, vendors, and system architectures.
- Focuses on future-ready intelligence standards featuring reproducible scorecards, capability tiering, and safety-critical thresholds for responsible real-world deployment.
MazeByte
SOTA Proof of Concept (PoC)Semi-Stealth
- Research project pioneering fully autonomous AI agents that self-manage end-to-end ETL pipelines without human intervention.
- Employs infant-inspired feedback mechanisms based on reward and pain signals, alongside resource-seeking and self-preservation behaviour to enable adaptive learning.
- Generates modular pseudo-code implementing memory, search, optimisation, and orchestration functionalities essential for continuous data pipeline automation.
- Built around the proprietary MBI internal model aiming to replace manual data engineering with resilient agent-driven automation.
SightBit
SaaS (TRL-9)Public
- Production-grade computer vision platform delivering real-time detection for drowning prediction, human and vessel identification, and flood-risk monitoring.
- Operates exclusively on standard CCTV video streams, eliminating the need for specialised sensors, edge computing devices, or per-site calibration.
- Incorporates a panoptic segmentation deep learning model enhanced with a proprietary augmentation pipeline, significantly improving small-object recall rates by 18 percentage points.
AdaptiveBridge
Production-ready (TRL-8)Public • Open-Source
- Open-source Python library designed to ensure machine learning model robustness against missing features during inference in real-world production.
- Predicts and imputes absent features dynamically, preserving model accuracy and stability.
- Supports configurable thresholds for feature importance, correlation, and accuracy metrics, alongside custom accuracy logic and automated feature distribution handling.
- Provides a drop-in estimator interface enabling seamless integration with existing ML pipelines via fit-transform workflows.
- Manages complex feature dependencies including mandatory, deviation-aware, and leveled feature types to optimise predictive performance.
Jhive
Production (TRL-7)Public • AIforGood
- Django-based web application leveraging OpenAI's API for real-time detection and analysis of antisemitic content.
- Implements robust content classification workflows paired with a detailed reporting interface for thorough review and operational transparency.
- Exposes RESTful API endpoints with integrated rate limiting to support scalable integration.
- Enforces strict input validation and sanitisation protocols to mitigate injection and security risks, ensuring safe, reliable deployment for sensitive content moderation scenarios.