PRODUCTION ML / INTELLIGENT SYSTEMS

Engineering
intelligent systems
at scale.

Senior Staff Software Engineer at Google, building production ML systems across reinforcement learning, retrieval, and computer vision. Now advancing neuromorphic computing.

View Experience Request résumé
Leto Hillza
ACTIVE

Engineering Approach

01. Production First

Start with the real problem. Build systems that meet latency budgets, adapt to changing data, and remain reliable under production traffic.

02. Efficient by Design

Treat compute, memory, and energy as design constraints, from large-scale model serving to event-driven inference at the edge.

03. Built to Evolve

Make learning systems observable and reproducible, so new models and architectures can improve the system without compromising reliability.

Selected Systems

Representative systems across reinforcement learning, retrieval, model delivery, and neuromorphic runtime design. The through-line is production ML: high-scale inference, efficient adaptation, and architectures that survive real traffic.

Reinforcement Learning

Budgeting Optimization System (BOS)

A real-time decision engine for Google Ads that used reinforcement learning to optimize pacing, allocation, and market liquidity across multi-billion-dollar spend.

Retrieval & Ranking

Semantic Discovery Stack

Retrieval and ranking systems for Google Drive and Ads that combined embeddings, nearest-neighbor search, and intent signals to surface the right result under tight latency budgets.

ML Platform

ML Deployment & Model Freshness

Production pipelines for training, deployment, and low-latency serving that kept models fresh in hours instead of days while maintaining reliability at global scale.

Neuromorphic Runtime

Hybrid SNN Edge Runtime

A low-power runtime for spiking neural networks that pushed event-driven inference closer to the sensor, reducing wasted compute and power on edge hardware.

Career Trajectory

Experience

Google Feb 2026 — Present

Senior Staff Software Engineer

Neuromorphic Computing & Hybrid AI

  • Integrating spiking neural networks (SNNs) into scalable production systems, developing spike-timing-dependent plasticity (STDP) learning rules for real-time adaptation.
  • Designing neuromorphic algorithms targeting sub-milliwatt inference on custom silicon — optimized for low-power edge devices.
  • Co-designing AI accelerators (TPU/ASIC) with hardware teams, translating biological neural dynamics into dataflow architectures.
Google 2022 — 2026

Staff Software Engineer

Applied ML & RL Infrastructure

  • Built reinforcement learning systems that optimized Google Ads budgeting and pacing for multi-billion-dollar spend in real time.
  • Designed online learning, retrieval, and inference pipelines that kept ads models fresh at trillions-of-events scale.
Google 2020 — 2022

Senior Software Engineer

Ranking, Retrieval & ML Platform

  • Built ML deployment systems for Google Workspace that cut model release latency by 60%.
  • Scaled collaboration and model-serving systems through a 300% traffic surge without degrading user experience.
Google 2018 — 2020

Software Engineer

Computer Vision & Multimodal ML

  • Built computer vision and retrieval systems for real-time visual understanding in Google Lens.
  • Scaled multimodal inference pipelines that powered visual discovery for 1B+ users.

Current Focus / Neuromorphic Computing

From production ML to
brain-inspired compute.

The next chapter builds on the same engineering foundation: making intelligent systems useful, efficient, and reliable. My current work brings spiking neural networks and hardware-aware design into that picture.

Read my neuromorphic writing (opens in a new tab) Explore the core concepts

Event-driven learning

Spiking neural networks and learning rules for real-time adaptation.

Efficient edge inference

Low-power runtimes that bring inference closer to sensors and devices.

Hardware & software co-design

Algorithms and dataflow architectures developed alongside accelerator hardware.

THE IDEAS BEHIND THE SYSTEMS

Core Concepts

Select a card to explore. Use Enter or Space with a keyboard.

SNN

SPIKING NEURAL NET

RL

REINFORCEMENT LEARNING

VON NEUMANN

THE BOTTLENECK

EDGE

LOCAL INFERENCE

Research & IP

PATENT US-110452A Pending

Event-Driven Memory Allocation for Sparse Tensors

REQUEST_DOCS
WHITEPAPER Google Internal

Optimizing RL Agents for TPUv4 Architecture

READ_WHITEPAPER

Let’s build what’s next.

Working on large-scale ML systems, neuromorphic computing, or production AI architecture? Let's talk.

Get in touch

System_Dependencies

Python
ML SYSTEMS
C++
LOW LATENCY
Go
BACKEND SYSTEMS
Rust
MEMORY SAFE
JAX
AUTOGRAD
Verilog
HDL
Torch
DEEP LEARNING
TPUv4
HARDWARE