Data platforms. AI systems. The infrastructure between them. Engineering Leader at Google Cloud. AI Systems Architect.
api-mcp-compiler turns OpenAPI, Swagger, and WSDL specifications into MCP tool surfaces with field-level provenance, human approval gates, and compiled-in refusals. Four pre-registered evaluations could not resolve whether semantic planning beats one tool per operation. Here is the design argument, the numbers, and what the evaluation harness caught anyway.
I wrote a 4-bit weight-only GEMM in pure Triton. The fast W4A16 kernels are all CUDA, so this one runs on NVIDIA and AMD. It beats cuBLAS FP16 by 1.1 to 1.3x in the decode regime, and the road there was mostly me being wrong about memory bandwidth.
Prompt-time activation defenses stop GCG and AutoDAN cold, then fail half the time against prefilling. New paper: why that blind spot is structural, and how a linear probe on the first generated tokens closes it - AUROC 0.97-1.00 across seven models, zero prefilling success when composed with steering.
Everyone assumes activation probes are too expensive to run in production. I ran the numbers on paper and I no longer believe it. Here is the argument, my predictions, and the harness I built to settle it.
What DeepSeek's DSpark and DeepSpec release actually changes for LLM inference: suffix decay, confidence-scheduled verification, a decision guide, and how to train a draft model for your own weights.
A loop always converges. That it reached a stable 'done' state tells you nothing about whether that state is correct, because the loop selects for whatever passes your check, not for what you meant. A practitioner's take on loop engineering: why convergence is not correctness, why the verifier is the part that actually decides, and what changes when loops run unattended over real production…
Embedding space is the new ad real estate. Mapping LLM ad auctions, RAG poisoning, GEO, and a framework for what comes next.
I wrote a fused Mixture-of-Experts dispatch kernel in pure Triton that beats Stanford's CUDA-optimized Megablocks at inference batch sizes, and runs on both NVIDIA and AMD GPUs without a single line of CUDA.
Three labs. Three different bets on how to catch models misbehaving. Each makes different assumptions about when models 'know' they're lying. Here's what works, what doesn't, and what happens when you combine them.
OpenAI trained GPT-5 to confess when it misbehaves. It works surprisingly well - except when the model doesn't know it's misbehaving. That's where activation probes come in.