AI InitiativeA connected body of work
Research, architecture and product building towards intelligence that develops through experience. Start with a question, then explore the work behind it.
AI Initiative research
What helps intelligence gain a new ability without losing an earlier one? Modular neural learning, poker experiments and mathematical approaches make that question concrete.
Research into reuse, retention and learning effort; results stay connected to the setting in which they were observed.
Professional contributions at Fifth9
Conduir connects coordination, model access, memory and review. Role Fabric examines how responsibility and handoffs remain clear as people and agents work together.
The architecture links what a person wants to achieve with the capabilities and decisions needed to pursue it.
Professional contributions at Fifth9
Manthan explores how ideas develop through different perspectives, criticism and combination. GraFOSa makes the relationships in a system available for investigation.
These are complementary contributions: discover new possibilities, understand the existing system and choose what deserves a closer look.
Independent creation, professional work and concepts
Games are the current route to market. Learning experiences, document intelligence and shared spaces show other ways the larger direction can become tangible.
Explore the experience, its design and the lessons from building it. Each case retains its own stage and attribution.
Research lineages and future directions
Enigma, alternative memory research and earlier cognitive-system work broaden the enquiry into adaptive intelligence, self-understanding and continuity.
The point of a research programme is to turn worthwhile questions into meaningful experiments, including learning when a hypothesis does not hold.
Explore how a capability works, why it is designed that way and what experience has taught us. The diagrams and principles keep the thinking visible.
Can intelligence acquire a new skill without starting again?
Reusable neural components, task contracts and selective-training experiments provide a practical basis for investigating learning cost and retention. The portable checkpoint separates repair and routing tasks; useful transfer and generalisation remain research questions.
What can intelligence learn when information is incomplete and choices have consequences?
Poker research combines a game engine, reference solvers, neural agents and comparative evaluations. Matiz adds a poker application with odds and outs. Together they support distinct investigations of game correctness, decisions under uncertainty and the experience of play.
How can each piece of work make the next one more useful?
Conduir brings work routing, model access, memory, review and governance into a layered architectural programme. Its service implementations and protected boundaries support a larger undertaking: connecting capabilities into dependable journeys from human purpose to useful outcomes.
What should an AI collaborator remember when the work continues?
Cortex provides persistent records and retrieval for relevant context, with scope checks and provenance. Its memory design distinguishes immediate context from longer-lived records, while benchmark work examines whether retrieval returns the reference evidence needed for a task.
What could a change affect beyond the code directly in view?
GraFOSa represents files, symbols and code relationships as a graph that agents can query. Impact analysis and isolated shadow graphs support investigation of proposed changes. Structural context informs engineering judgement; runtime behaviour still needs execution and testing.
Can we preserve how an idea develops, including the disagreements that improve it?
Manthan records contributions from AI personas in a lineage graph, linking ideas that derive from, contradict, refine or combine others. A viewer and graph export make that development inspectable; the value of generated ideas requires separate evaluation.
How can different AI perspectives help us examine a piece of work?
The AI Review Panel assembles models from different vendors, then challenges, deduplicates and ranks their findings. It records which vendor families participated. Proposed concerns are checked against the work, so agreement between models is not treated as correctness.
How do multiple AI contributors become a team whose work can be understood?
The AI Engineering Workforce is a delivery practice that applies coordinated agents to software engineering, separating implementation from review and tying completion decisions to supporting evidence. Leased workspaces, admission limits and automated checks help organise concurrent contributions while preserving human judgement and responsibility.
What changes when people and AI can work within the same shared space?
Conduir Spatial combines a browser-based 3D interface, shared scene state and a physics engine. Signed operations and a policy gate distinguish proposed changes from accepted actions, providing foundations for attributable collaboration in a shared environment.
How can more ideas become experiences people can actually enter?
Game Creation Studio is a toolkit that combines reusable starting templates with shared foundations for saves, input and layout. AI-assisted implementation and review help turn ideas into playable experiments, while browser checks and release preparation support the path towards a usable game.
How might intelligence learn to organise itself?
Explore adaptive organisation, model coordination and the distinction between measurable behaviour and deeper questions about consciousness.
The portfolio spans research, implementation and products at different stages. The shared direction connects their purpose; each project explains its own contribution.