Projects

MHIL

Concept & Early Prototyping (TRL-5)
Semi-Stealth • Non‑Profit
  • MHIL is a non-profit research institute developing rigorous benchmarks and evaluation frameworks for genuine machine intelligence beyond narrow LLM tasks.
  • It constructs human-aligned test suites assessing agent capabilities such as memory retention, strategic planning, resilience, and autonomous problem-solving under uncertainty.
  • MHIL delivers an open evaluation matrix tailored for "pre-superintelligent" AI systems, enabling standardized, vendor-agnostic comparisons across architectures.
  • The lab targets future-proof standards including reproducible scorecards, capability tiers, and safety-critical thresholds essential for real-world deployment scenarios.

MazeByte

SOTA Proof of Concept (PoC)
Semi-Stealth
  • MazeByte pioneers fully autonomous AI agents designed to self-manage ETL pipelines without human intervention.
  • It implements infant-inspired feedback mechanisms using reward and pain signals alongside resource-seeking behaviour to drive adaptive learning and self-preservation.
  • The system autonomously generates pseudo-code modules handling memory, search, optimisation, and the entire ETL lifecycle.
  • Powered by the proprietary MBI model, MazeByte aims to replace manual data engineering by delivering end-to-end automation of complex data pipelines.

SightBit

SaaS (TRL-9)
Public
  • SightBit is a production-grade computer vision SaaS platform providing real-time detection of drowning risks, humans, vessels, and flood hazards using standard CCTV inputs.
  • The solution operates without specialised sensors, edge hardware, or site-specific tuning.
  • Its proprietary panoptic segmentation model incorporates a sophisticated augmentation pipeline that significantly elevates small-object recall performance, enhancing reliability in critical safety environments.

AdaptiveBridge

Production-ready (TRL-8)
Public • Open-Source
  • AdaptiveBridge is an open-source Python library designed to maintain machine learning model robustness by predicting and imputing missing features at inference time.
  • It supports configurable thresholds for feature importance, correlation metrics, and incorporates customizable accuracy functions alongside automated data distribution management.
  • Offering a drop-in estimator interface with fit and transform methods, it integrates seamlessly into existing ML pipelines.
  • The library intelligently manages feature dependencies—including mandatory, deviation, and leveled features—to preserve predictive performance.
  • Comprehensive documentation, unit testing, and continuous integration accompany its production-ready release under a permissive open-source license.

Jhive

Production (TRL-7)
Public • AIforGood
  • Jhive is a Django-based real-time web application utilising the OpenAI API to detect and analyse antisemitic content with precision.
  • It features detailed reporting interfaces for content review and offers RESTful API endpoints with robust rate limiting for secure integrations.
  • The system enforces stringent input validation and sanitisation to mitigate injection attacks
  • focusing exclusively on antisemitism detection rather than general toxicity or surveillance.