Kulbir Minhas connects long-term AI vision with the architecture, engineering and product work needed to make it useful. His work brings together learning systems, memory, agent collaboration, research and interactive products. AI Initiative is the independent lab through which he pursues a sustained interest in intelligence that can grow alongside people. His professional architecture work at Fifth9 forms a separate, clearly attributed part of this wider portfolio.
The evolution of humans and AI
“Humans and AI, growing in harmony.” Like voices in harmony, each stays distinct, and together they create what neither could alone. This is the belief behind the work.
AI is evolving in three steps: tools that waited for instructions and did one task; assistants that respond, draft and suggest, then forget; and, next, peers that remember, learn from outcomes and grow with the people they work with.
The partnership works when each side brings what it does best. People bring purpose, values and judgement; AI brings memory, reach and tireless checking. Every outcome teaches both: people decide better, and AI learns what matters. This is co-evolution: human and machine capability rising together, with success measured by what people can now do.
The vision: intelligence that grows with us
AI is advancing towards a future in which it could become a continuing peer: a collaborator that understands shared work, contributes different abilities and learns through experience. AI Initiative explores how to make that relationship useful, accessible and accountable, while helping people develop their own understanding and capabilities.
The ambition is an evolving AI system in which learning, memory, exploration and execution reinforce one another. Useful progress should carry forward between tasks, tools and models. A correction, a successful method or a failed experiment can inform what happens next. The selected work represents current contributions to this wider destination, not its complete capability set or a single fully implemented system.
How the capabilities connect
Compound learning combines multiple language models with specialist neural networks, mathematical methods, tools and human judgement. Improvement can happen through neural training, but also through better routing, reusable procedures, team composition and evaluation while the underlying models remain unchanged. The system should learn which capability to use, how to combine it and how to judge the result.
Memory provides continuity: retaining useful context with provenance, accommodating correction and forgetting, and respecting personal and shared boundaries. The longer-term aim is continuity independent of any one language model. In the intended cycle, execution puts that knowledge to work and captures outcomes, with evaluated learning informing future decisions. Exploration develops alternative ideas; self-improvement tests better ways of using existing components before proposing new ones. The benefit of connecting these capabilities must be demonstrated in real tasks.
Selected work and its contribution
At Fifth9, the architectural work includes Conduir and Role Fabric for coordinating capabilities and collaborative execution; Cortex for memory and continuity; GraFOSa for structural understanding through graphs; and Manthan for sustained exploration, critique and idea development. These are complementary areas of work with their own implementation paths. They should not be read as interchangeable product names or as a statement that all intended capabilities are deployed.
Independent work through AI Initiative includes modular neural learning, poker agents and the Matiz poker app (odds and outs), and interactive learning games built through the Game Creation Studio. These provide practical settings for investigating decisions, reuse, feedback and human–AI interaction. Games are a current route to market and to broader participation, with other applications remaining open.
The wider research portfolio also explores quantum-inspired representations, associative memory and Enigma’s questions about adaptive organisation and emergence. These are investigations, not claims of demonstrated quantum advantage or machine consciousness. DocuMind is a proposed document-intelligence application connecting source material, exploration, human review and useful outputs.
What the research is teaching
Learning should be assessed through retained capability and useful outcomes as well as speed. Selected neural-transfer experiments have explored whether reusing a learned component can reduce training on a new task while preserving an earlier one. The findings are task-specific: faster adaptation does not establish lower total cost once earlier training and retention checks are counted.
Failures matter too. Interpolation may work where extrapolation does not; adapting a model can damage earlier knowledge; and an automated check can miss a problem visible in real interaction. These lessons inform the broader architecture: preserve context, test transfer, make evaluation explicit and keep human judgement in the learning loop.
A journey from enterprise systems to evolving AI
Kulbir is Principal Data Scientist at Fifth9, where he is based at HSBC in Developer Experience and architects Conduir, an agentic, multi-agent AI architecture in which people and AI agents work together, pointing towards AI as a continuing peer that grows with people. He founded AI Initiative in 2024. Previously, he co-founded Insight Delivered and progressed from Director R&D to CTO, combining technical direction, team leadership and AI/ML product development, including AutoML, DeX and Strategic Finance. Earlier financial-services and data experience includes work with Morgan Stanley and Credit Suisse.
The path ahead
The path Kulbir is taking: from AI that answers to AI that grows with people, and from building systems to shaping how humans and AI evolve together.