From Blockchain Research to LLMs

In 2022, I published a paper on using permissioned blockchains for electronic health record management. The system combined Hyperledger Fabric, IoT sensors, and predictive AI to create a decentralized, privacy-preserving health data platform.

That work taught me something that carries directly into my current focus on LLMs and agentic AI: the hardest problems aren't in the model — they're in the system design around the model.

The parallels

Blockchain and LLM systems share a surprising amount of design philosophy:

• Trust boundaries: In blockchain, you design around untrusted participants. In LLM systems, you design around non-deterministic outputs. Both require explicit verification layers.
• Consensus vs. validation: Blockchain nodes reach consensus on state. LLM pipelines validate outputs against schemas and constraints. The mechanism differs, but the goal is the same — reliable system behavior from unreliable components.
• Data provenance: Health records need audit trails. AI systems need observability. Both are about knowing what happened and why.

What changed

The biggest shift from blockchain to LLMs was speed of iteration. Blockchain systems are slow to deploy and hard to change — that's by design. LLM-based systems can be iterated on rapidly, but that speed creates its own challenges: prompt drift, regression, and the temptation to ship without adequate evaluation.

Research taught me to be rigorous about evaluation methodology. That discipline is valuable when building with LLMs, where it's easy to mistake "it looks right" for "it works correctly."

What stayed the same

The fundamentals haven't changed: understand the problem deeply before picking the tool, design for failure modes you haven't seen yet, and always measure what matters.