Hi, I'm Omoshola.
AI/ML researcher and systems builder. I work on the seam between what AI can do and what institutions can trust. The work spans three layers that have to hold together: specialist models that go deep in a single domain instead of wide across many, the infrastructure that gives those models lasting memory and a verifiable identity, and a settlement layer that turns every decision involving value into a record anyone can independently replay.
I work on the seam between what AI can do and what institutions can trust. Day to day that means supply chain and financial systems: I work as a Kinaxis supply-chain subject-matter expert, building AI-enhanced predictive planning and supplier-risk intelligence for operations where a wrong recommendation has real cost. The interesting questions in these systems have stopped being about capability. The open ones are about accountability. Whether an autonomous decision can be replayed. Whether a model's memory belongs to the user or the vendor. Whether a regulator can independently verify what happened, three years after it happened, without the original operator in the room. The work I do spans three layers that have to hold together: specialist models that go deep in a single domain instead of wide across many, the infrastructure that gives those models lasting memory and a verifiable identity, and a settlement layer that turns every decision involving value into a record anyone can independently replay.
Most of what I build, write, and review lives in that gap. The writing is about what the gap actually looks like in code; the work is the long version.
Recent Posts
-
Too Central to Fail: Finding the Suppliers That Can Take Down a Network
Published:• 8 min readThe supplier that fails and takes your whole operation with it is rarely your biggest line of spend. It is the one buried three tiers deep that forty other parts quietly depend on. Notes on network analysis, cascading disruption, and building supplier-risk scores an operator can actually act on.
-
Why Supply Chain AI Fails at the Last Mile
Published:• 9 min readThe hardest part of supply chain AI is not the forecast. It is the handoff — the moment a probabilistic recommendation reaches a human planner who has to decide whether to trust it. Notes on the trust gap between a recommendation and an action, from the seat of a Kinaxis subject-matter expert.
-
Verity: The Truth Engine Underneath XAP
Published:• 14 min readA close read of verity-engine — five Rust crates, seven trust properties, and the design invariants that turn an agent settlement decision into a record any third party can independently replay.
-
XAP: The Settlement Layer for Autonomous Agents
Published:• 10 min readA close read of XAP (eXchange Agent Protocol): six primitive objects, one transaction flow, and what it takes to make agent-to-agent commerce verifiable without any human in the loop.
-
AgenticMemory, Part 3: The Capabilities That Required a New Data Structure
Published:• 9 min readFour things that are only possible because of the typed cognitive graph — finding connections, simulating what breaks, auditing gaps in reasoning, and recognizing patterns across completely different domains.
-
AgenticMemory, Part 2: How the Agent Searches and Knows What Matters
Published:• 8 min readSelf-correction without data loss, exact-term search that actually works, hybrid retrieval that uses both signals, and a way to find which beliefs everything else is built on.
-
AgenticMemory, Part 1: The Memory Your Agent Was Always Missing
Published:• 10 min readFour foundational capabilities that give AI agents a real brain — not a search bar. The binary format, the portable file, the typed events, and the reasoning chains.
-
Building a Payment System That Proves Finality Instead of Asserting It
Published:• 11 min readMost payment platforms tell you a transaction succeeded. I am building one that proves it — cryptographically, deterministically, and in a way that can be replayed and audited years later. Notes from the work in progress.
-
Building a Memory System for AI Agents
Published:• 22 min readVector databases tell you what is similar. They do not tell you what happened, what was decided, or what the agent learned that overrides what it knew before. I needed something different.
-
Building a Web Cartographer for AI Agents
Published:• 21 min readI wanted AI agents to understand the web the way a researcher does — not just fetch a page, but navigate structure, follow intention, and remember where they have been. Notes from building Cortex.