Luca Lowndes
Email me
Founding engineer · AI systems · HPC research

I build ambitious
AI products and the
systems behind them.

I’m Luca Lowndes, founding and lead engineer at EvryStay. I built the product as its only engineer, helped take it through a A$4.5M raise, and now lead a team of five while continuing to ship.

I’m also a first author and speaker at ICPP 2026 for work on serving DeepSeek-R1 671B with SGLang, a credited SGLang contributor, and the youngest UniHack winner on record.

Explore my work Read the paper
Founding engineer ICPP 2026 first author SGLang contributor UniHack winner APAC HPC-AI 3rd place Founding engineer ICPP 2026 first author SGLang contributor UniHack winner APAC HPC-AI 3rd place
01

Experience

From zero to product,
then product to team.

I work closest to the hard boundary between research, product, and production: turning an idea into something people can depend on.

Dec 2025 — Present

Mornington, VIC

EvryStay

Founding and Lead Engineer

Current

Built the company’s product from zero as its first and only engineer, then built the engineering team around it.

Built the product

Designed and built Eve, an AI concierge that handles guest conversations end to end for short-term rental hosts, property managers, and hotels, together with the PMS integrations and AI service architecture underneath it.

Backed by investors

The platform I built was the product investors backed. EvryStay opened and closed a A$4.5M round at a A$17M post-money valuation inside one week.

Grew the team

Hired and now lead five full-time engineers in person. I own architecture, the AI service boundary, code review, and final technical calls while still shipping daily.

5 engineers led 300+ merged PRs A$4.5M raised

Mar 2025 — Nov 2025

Melbourne, VIC

Monash DeepNeuron & Monash eResearch

Lead AI Researcher, APAC HPC-AI Competition Team

Optimised distributed inference for DeepSeek-R1 across two 8×H200 nodes using infrastructure from NSCC Singapore, NCI Australia, and Firmus.

The finding

A deliberately configured single node beat a default 16-GPU setup: PP2 · TP4 · DP4 attention reached 17,417 tok/s and 2.05× the per-GPU efficiency.

The lesson

The “boring” systems work mattered most. CUDA version alignment alone delivered +35%, while more than 40 hand-tuned NCCL configurations never beat the defaults.

Outcome Presented at SupercomputingAsia 2026 in Osaka; the work became the ICPP paper.

Feb 2025 — Present

Melbourne, VIC

Monash DeepNeuron

AI & HPC Technical Team, Outreach

Researching Neural Cellular Automata for self-organising biological systems and running practical AI workshops in Victorian high schools.

Jan 2025 — Nov 2025

Hybrid

Scag Australia · International Mowers

Freelance AI Engineer

Built customer-facing AI agents for model comparison, troubleshooting, and service recommendations, with structured proprietary data, embeddable web/mobile chat, brand persona, and legal guardrails.

02

Research

Less is more.

Our results challenge the instinct to distribute the biggest model across every available GPU. Topology-aware parallelism was both faster and more efficient.

ICPP 2026 · BID ’26 First author

Less is More: Optimising SGLang Distributed DeepSeek-R1 Inference on a Two-Node H200 Cluster

Luca Lowndes · Nathan Culshaw

Accepted at Benchmarking in the Data Center, a workshop of the 55th International Conference on Parallel Processing. To appear in the ACM Digital Library.

Read PDF Experiment record
+81% throughput over
16-GPU baseline
Best result
17,417 tok/s
Hardware
8× NVIDIA H200
Model
DeepSeek-R1 671B
Serving stack
SGLang
Core insight

Keep collectives on NVLink and replicate within a node instead of sharding critical communication across InfiniBand.

01

Talk · Singapore · Sep 2026

Presenting “Less is More” at BID ’26

Speaking alongside representatives from NVIDIA and the HPC-AI Advisory Council.

02

Panel · ICPP 2026

Accelerating the Future

Panellist on overcoming bottlenecks in GPU-based AI and HPC, with NVIDIA and the HPC-AI Advisory Council.

03

Talk · Osaka · Jan 2026

SupercomputingAsia 2026

Presented the APAC HPC-AI Competition work after placing third among 49 university teams.

03

Selected projects

Useful things,
built all the way through.

Agents that write production-shaped code, models that grow biological structure, and experiment records designed to survive scrutiny.

2026 Company agent

Dummy

Describe a feature in Slack. Dummy builds a working version.

It creates real branches across three repos, a seeded full-stack preview at a permanent URL, desktop and mobile QA screenshots, and draft pull requests.

Slack thread
→
Plan + build
→
Review agent
→
Preview + PRs

Why: The non-technical team can see ideas working before engineers commit time to them.

Guardrails: Independent code review, approval from an allowlisted human for sensitive changes, and no access to merge, deploy, or production. Built on the open-source QM harness.

2025 Hackathon winner

Growth Garden

Habit tracking as a garden you grow by getting things done.

On a team of six, I architected the contextual AI assistant in Langflow, built the DataStax model and live sync, and led the winning pitch.

1st of 135+ projects
2025 — Present Research

Neural Cellular Automata

Learning complex growth from simple local rules.

Built the 3D temporal training path for time-resolved mouse embryo data to explore gastrulation-like development, including reduced voxel handling, dynamic frame batches, differentiable IoU, loss instrumentation, PyTorch models, and browser inference work.

Visit neuralca.org
2025 Open experiment record

APAC-AI-HPC-2025

The reproducible record behind every number in the ICPP paper.

Launch scripts, Slurm logs, configuration notes, and workbooks for every experiment across two 8×H200 nodes.

github.com/LucaLow/APAC-AI-HPC-2025 Open repository
04

Recognition & education

A fast start,
not a finished story.

I deferred university in 2026 to build EvryStay properly, while continuing research, open-source work, and conference speaking.

Awards

04
03

APAC HPC-AI Competition

3rd of 49 university teams · Presented at SupercomputingAsia, Osaka

01

UniHack 2025

1st of 850+ participants · Youngest winner on record

HD

Engineering Dean’s Award

Academic Excellence · Top of the FIT1045 cohort, Monash

D

Dux of Mathematical Methods

Elwood College

Education

02

Monash University

BEng (Hons) Mechatronics / BCS Algorithms & Software

  • High Distinction WAM
  • Deferred in 2026 for EvryStay

Academy of Interactive Entertainment

Certificate III in Information Technology

  • Completed alongside VCE
05

Technical toolkit

Deep enough for the kernel.
Broad enough for the product.

01

Inference & HPC

SGLang, CUDA, NCCL, tensor parallelism, pipeline parallelism, data parallelism, Slurm, Singularity, NVIDIA H100/H200

02

Machine learning

PyTorch, TensorFlow, neural cellular automata, distributed model serving, evaluation, loss design

03

Product engineering

TypeScript, Next.js, Python, Postgres, LLM agents, tool use, multi-repo systems, production architecture

04

Systems & delivery

C, C++, Git, CI/CD, code review, technical leadership, hiring, architecture ownership

Let’s talk

Building something
worth obsessing over?

Research, product, distributed AI systems, or the point where all three collide.

Luca@lowndes.net