How It Started
It began with frustration — not inspiration.
Ravi Natarajan had spent more than three decades building enterprise systems — COBOL and C, then Grady Booch's Object Oriented Design with Applications and OOA/OOD, into C++, UML, Java, distributed systems, SOA, design patterns, TOGAF. Decades of learning that structure matters, that design precedes code. So when agentic AI took off — teams reaching straight for LLMs, asking agents to reason, design, code, and orchestrate everything at once, with no model, no boundaries, none of the foundational principles the craft had spent decades establishing — it troubled him deeply. The craft was being bypassed entirely.
That is what prompted him to stop, think, and start capturing his own thinking on how this should be built — properly — before writing a line of code. In early 2025 that became practice: recorded on a voice recorder on his iPad, architecture as spoken word — patterns, boundaries, contracts, governance, orchestration — before a single line of code was written.
Those recordings captured the thinking that became K9-AIF's foundation: base classes, Architecture Building Blocks (ABB), Solution Building Blocks (SBB), and the contracts between them. The roots went back further — a few years earlier Ravi had been trying out his old C++ design patterns in Python, which became an LLM Factory, then a Retriever component: disciplined, purposeful, trained to find exactly what's needed. That is where K9 came from. Drawing on his TOGAF background, he applied that same thinking to ABB and SBB — and that is what evolved into the framework.
That is what makes K9-AIF different. The architecture isn't just in the code — it's encoded in a form AI itself can understand and apply correctly. The framework evolves; the architectural intent does not drift.
That claim has since been tested in practice: working from CLAUDE.md
and SKILLS.md, Claude Code generates accurate code that complies with
the ABB contracts on its own — efficiently, and without drifting from the
architecture. With that strong a foundation in place, further development only gets
easier from here.
Journey
That same rigor — not a methodology borrowed from a textbook, but decades of structuring complexity, defining boundaries, and governing behavior — is what K9-AIF is built on. Abstraction, inheritance, dynamic binding, polymorphism — the core OOA/OOD principles from the Booch, C++, and Patterns era — are applied throughout, from the ABB/SBB contracts down to the way every agent is defined and extended. Applied to agentic AI, it produced something the industry has been missing: a framework that treats agents as first-class architectural concerns, not just scripts with an LLM attached.
That conviction is laid out at length in
Boxer —
an essay that takes BMW's Big Boxer engine, recognizable since 1923 through a century
of technological change, as the model for what K9-AIF aims to be: "Technologies
shift. Architecture shouldn't have to." Models, providers, and infrastructure can
be swapped out; the exposed, inspectable ABB/SBB contracts —
BaseAgent, K9EventRouter, BaseModelRouter —
are built to stay.
K9-AIF is open source — organizations are free to take it, customize it, and extend it as suits their needs. Ravi's wish is simple: that it keeps growing, that it expands, and that real applications get built on it.