Nik Liolios

Personalization for AI Mode — Google's generative search experience. The systems that tailor AI-generated answers to a user's context, history, and intent.

Personal Intelligence in AI Mode ↗

Led the launch of the Images feed — the ranked, browsable feed surface in Google Images. Ranking model through to the production surface.

Explore and organize images ↗

Retrieval and ranking for Google Discover — the recommendation feed in the Google app and on Android. The models behind what gets surfaced, and the infrastructure to serve them.

Before Google — recommendation and ranking for the Instagram feed. Retrieval and ranking models for the main feed.

A GPT-style transformer written and trained from scratch on a 16 GB Mac mini — no CUDA, no PyTorch. To make it fit and run on the M3 GPU through the open IREE stack, I hand-wrote the hot Metal kernels: a native FlashAttention (forward and an O(seq)-memory backward), a bf16 GEMM, and a fused cross-entropy. A toy model, but a real look at training on constrained Apple hardware.

Apple abandoned jax-metal, so there was no open, current way to run JAX on an Apple GPU properly. This is a fork of the IREE compiler that teaches it to run bf16 matmuls on the GPU's matrix units instead of scalar loops — roughly doubling throughput, correct across a suite of standard transformer and vision models.

A reverse Turing test: finish a few sentences and it measures the surprisal — how predictable each word you write is to a panel of language models. The less you surprise the machines, the more clanker you score.

Former founder — built Tetra (YC W17; AI phone-call transcription, acquired in 2018). Based in NYC. Outside work I race triathlon — most recently IRONMAN 70.3 New York.