I build backend systems
for AI infrastructure.

Most AI systems don't fail because the model is weak. They fail because nobody can see what happened between retrieval, prompting, tool execution, and state updates.

Memory is becoming more important than raw model size. A smaller model with stable retrieval and consistent state management is usually more useful than a larger one that forgets why it made a decision three steps ago.

Most multi-agent systems eventually rediscover distributed systems problems: retries, consensus, partial failure, stale state, and conflicting writes. The hard part is coordination, not generation.

A surprising amount of 'trust' in AI systems is just good UI. If verification depends on the interface instead of independently reproducible traces, you're still relying on vibes.

Retrieval quality changes with model scale. In our smaller VLM experiments, retrieval did most of the work. As models improved, visual grounding started mattering more than the retrieved examples themselves.

Most on-chain AI projects don't actually need a blockchain. The interesting use cases start when multiple parties need shared state, coordination, or verification without trusting a single executor.

Most coding-agent demos assume the environment behaves perfectly. Real systems need retries, bounded execution, isolation, logging, and ways to recover when the agent inevitably does something stupid.