Portfolio — Applied AI · Security · Infrastructure
Software engineering leader & technical founder
I build applied AI systems — LLM-powered products, agent infrastructure, and the security and performance layers underneath. My work falls into three areas below — six projects are the core of it, and the full repository catalog has the rest.
Making AI cheaper and faster to run — the economics of LLM inference, GPU workload infrastructure, and CUDA performance engineering where it counts.
ProblemKernel benchmarks report GPU wins that vanish the moment you measure the whole application, because nobody accounts for PCIe data movement.
ProblemLLM prompt caches are prefix-sensitive, so tiny changes such as timestamps, UUIDs, reordered tools, or volatile context can invalidate thousands of reusable tokens.
Evidence over vibes — authorization with provenance, adjudication that tries to falsify itself, and reliability layers for agent systems.
ProblemAn agent's signature proves who asked for an action, not that the exact action was independently checked, unchanged since approval, or never executed before.
ProblemMultiple LLMs can agree on a convincing conclusion while sharing the same unsupported assumption, and agreement is not independent evidence.
Quantum protocols, error correction, and benchmarks run against real hardware — including honest null results.
ProblemAgents are already buying ads, booking travel, and moving money on behalf of humans over messages protected by cryptography that a quantum computer will break, while harvest-now-decrypt-later adversaries record everything today.
ProblemTensor-network claims for noisy quantum simulation conflate unique-record harvesting with real execution speedup, and nobody had measured exactly where batching actually wins.
Software engineering leader and technical founder building applied AI — from LLM cost optimization and agent reliability to post-quantum protocols. Based in Milpitas, California.
Building in AI infrastructure, security, or quantum?
Let's talk →