Projects
MHIL
Concept & Early Prototyping (TRL-5)Semi-Stealth • Non‑Profit
- MHIL is a non-profit research institute pioneering comprehensive benchmarks for authentic machine intelligence that extend beyond narrow LLM capabilities.
- It develops human-aligned test batteries designed for composite AI systems and autonomous agents, rigorously evaluating memory, planning accuracy, robustness, and problem-solving under uncertainty.
- The lab crafts open evaluation matrices enabling objective, vendor-agnostic comparisons across models and architectures, addressing the emerging demand for pre-superintelligent system assessment.
- Its frameworks focus on standardising reproducible scorecards, defining capability tiers, and establishing safety-critical thresholds vital for real-world, safety-sensitive AI deployment.
MazeByte
SOTA Proof of Concept (PoC)Semi-Stealth
- MazeByte is an advanced research initiative demonstrating a fully autonomous ETL AI agent capable of self-maintenance without human intervention.
- It employs biologically inspired feedback mechanisms, including reward and pain signals, alongside resource-seeking behaviours to foster adaptive learning and system resilience.
- The agent generates modular pseudo-code components that manage memory
- search
SightBit
SaaS (TRL-9)Public
- SightBit is a production-grade computer vision SaaS platform delivering real-time detection for drowning prevention, human and vessel identification, and flood-risk forecasting.
- Operating exclusively on standard CCTV video streams, it requires no specialised sensors or edge processing hardware and eschews site-specific calibration.
- Its core panoptic segmentation model, enhanced by a proprietary data augmentation pipeline, achieves significant improvements in small-object recall, offering scalable, reliable visual analytics for safety-critical environments.
AdaptiveBridge
Production-ready (TRL-8)Public • Open-Source
- AdaptiveBridge is an open-source, production-ready Python library engineered to maintain machine learning model robustness amid missing features at inference time.
- It predicts and imputes absent inputs with configurable thresholds on feature importance and correlation, supporting custom accuracy metrics and automated handling of variable distributions.
- Comprehensive documentation, thorough unit testing, and continuous integration accompany its permissive open-source release.
- Delivered as a drop-in estimator with a familiar fit-transform interface
- it manages complex feature dependencies
Jhive
Production (TRL-7)Public • AIforGood
- Jhive is a Django-based, production-level web application utilising the OpenAI API for real-time antisemitic content detection.
- It features a detailed reporting interface for monitoring classified content and provides RESTful API endpoints with rate limiting for secure integration.
- The system incorporates rigorous input validation and sanitisation protocols to defend against injection attacks
- prioritising operational security and responsible AI deployment in sensitive content moderation domains.