One prompt

Models

Prompt & details

Make me a personal website given all this info: [my info]

The info is a fixed snapshot of my experience, projects, research and education. There are no screenshots, styles or source files from this site in the brief.

Full prompt

Make me a personal website given all this info: Luca Lowndes
Founding and lead engineer at EvryStay, where I built the product as the only engineer, helped take it through a A$4.5M raise at a A$17M post-money valuation, and now lead a team of five. First author and speaker at ICPP 2026 on serving DeepSeek-R1 (671B) with SGLang, work that grew out of placing 3rd in the APAC HPC-AI Competition. Credited SGLang contributor. Youngest ever UniHack winner. Deferred my double degree at Monash to do this properly.

Luca@lowndes.net: mailto:Luca@lowndes.net
LinkedIn: https://www.linkedin.com/in/luca-lowndes/
GitHub: https://github.com/LucaLow

Experience
EvryStay. Founding and Lead Engineer. Dec 2025 to Present. Mornington, VIC
Built the product from zero. Joined as the first and only engineer. Designed and built Eve, an AI concierge that handles guest conversations end to end for short-term rental hosts, property managers and hotels, along with the property management system integrations and AI service architecture underneath it.
Fundraise. The platform I built was the product investors backed: EvryStay raised A$4.5M at a A$17M post-money valuation, opening and closing the round inside a week.
Grew the team. Hired and now lead five full-time engineers in person. I own the architecture, the AI service boundary, code review and the final technical call, while still shipping daily: 300+ merged PRs in my first seven months.
Monash DeepNeuron & Monash eResearch. Lead AI Researcher, APAC HPC-AI Competition Team. Mar 2025 to Nov 2025. Melbourne, VIC
The problem. Maximise inference throughput for DeepSeek-R1 (671B parameters, 37B active) across two 8×H200 nodes, using infrastructure from NSCC Singapore, NCI Australia and Firmus.
What I found. Fewer GPUs arranged deliberately beat more GPUs arranged by default: a tuned single node (PP2 · TP4 · DP4 attention) hit 17,417 tok/s at 2.05× the per-GPU efficiency of 16 GPUs. The boring stuff mattered most: CUDA version alignment alone was worth +35%, and 40+ hand-tuned NCCL configs never beat the defaults.
Outcome. Presented on stage at SupercomputingAsia 2026 in Osaka. The work became the ICPP paper.
Monash DeepNeuron. AI & HPC Technical Team, Outreach. Feb 2025 to Present. Melbourne, VIC
Research. Neural Cellular Automata for modelling self-organising biological systems (see Selected Projects).
Outreach. Ran hands-on workshops in Victorian high schools on how deep learning models work and how to use generative AI well.
Freelance AI Engineer. Scag Australia, International Mowers. Jan 2025 to Nov 2025. Hybrid
Customer-facing AI agents. Built the Scag Mechanic and IM Advisor assistants for model comparison, troubleshooting and service recommendations. Structured proprietary product data, shipped embeddable chat widgets for web and mobile, and put brand persona and legal guardrails around both.

Research
Less is More: Optimising SGLang Distributed DeepSeek-R1 Inference on a Two-Node H200 Cluster. Paper, first author with Nathan Culshaw. ICPP 2026
Venue. Accepted at Benchmarking in the Data Center (BID ’26), a workshop of the 55th International Conference on Parallel Processing. To appear in the ACM Digital Library.
Result. Keeping collectives on NVLink and replicating within a node beats sharding across InfiniBand: 17,417 tok/s, 81% over the 16-GPU baseline.
https://github.com/LucaLow/APAC-AI-HPC-2025/blob/main/LessIsMore.pdf
Presenting Less is More at BID ’26. Talk. Singapore, September 2026
Session. Alongside speakers from NVIDIA and the HPC-AI Advisory Council.
Accelerating the Future: Overcoming Bottlenecks in GPU-Based AI and HPC. Panellist, with NVIDIA and the HPC-AI Advisory Council. ICPP 2026

Projects
Dummy. EvryStay’s company agent, lives in Slack. 2026
What it does. Describe a feature in a Slack thread and Dummy builds a working version: real branches across three repos, a seeded full-stack preview at a permanent URL, desktop and mobile QA screenshots, and draft PRs.
Why. So the non-technical side of the company can see their ideas working before engineers spend time on them.
Guardrails. An independent review agent checks the code, sensitive changes need an allowlisted human to approve in Slack, and nothing it does can merge, deploy or touch production. Built on the open source QM harness.
Growth Garden. UniHack 2025 winner, team of six. 2025
What it is. Habit tracking as a garden you grow by getting things done. 1st of 135+ projects at Australia’s largest student hackathon.
My part. Architected the AI assistant in Langflow (contextual chaining, no fine-tuning), built the DataStax data model and live sync, and led the pitch to the judges.
Neural Cellular Automata. Monash DeepNeuron research, live at neuralca.org. 2025 to Present
What it is. Models that grow complex patterns from local rules, including a 3D NCA trained on time-resolved mouse embryo data to explore gastrulation-like development.
What I built. The 3D temporal training path—reduced data and voxel handling, dynamic frame-to-frame batches, differentiable IoU and loss instrumentation—plus PyTorch models and browser inference work.
APAC-AI-HPC-2025. Open experiment record behind the ICPP paper. 2025
What’s in it. Launch scripts, Slurm logs and workbooks for every configuration tried across two 8×H200 nodes.
Why it matters. Every number in the paper can be reproduced from it.

Awards
APAC HPC-AI Competition. 2025 to 2026. 3rd of 49 university teams, presented at SupercomputingAsia, Osaka
UniHack 2025. 2025. 1st of 850+ participants, youngest winner on record
Engineering Dean’s Award for Academic Excellence. 2025. Monash, top of the FIT1045 cohort
Dux of Mathematical Methods. 2024. Elwood College

Education
Monash University. 2025 to Present. Melbourne, VIC. BEng (Hons) Mechatronics / BCS Algorithms & Software. HD WAM. Deferred 2026 for EvryStay.
Academy of Interactive Entertainment. 2023 to 2024. Melbourne, VIC. Certificate III in Information Technology, completed alongside VCE

Skills
Inference & HPC. SGLang, CUDA, NCCL, tensor/pipeline/data parallelism, Slurm, Singularity, NVIDIA H100/H200
ML. PyTorch, TensorFlow
Product. TypeScript, Next.js, Python, Postgres, LLM agents and tool use
Other. C, C++, Git, CI/CD

Frozen 2026-09-06

Method

One session. One initial prompt per attempt. Each model runs in its chosen harness or chat app and can use its normal tools. There is no imposed time limit.
Same start. Coding runs start in a fresh folder containing the prompt. Chat runs use a new conversation. There is no starter app, required framework or supplied design system.
Every attempt. Fresh folders are available for up to 3 attempts per pairing. Repeats are optional. Failed attempts stay in the record.
The harness. The chooser labels the model and harness together. This compares the complete coding setups, including differences in their tools and instructions. Imported archives are labelled when the original prompt or transcript has not been supplied.
Same capture. Desktop screenshots use a 1440 × 900 viewport. Mobile uses 390 × 844. Live websites fit the browser window. Screenshots use local assets only, with external requests blocked and motion reduced.
The measurements. Time, reported usage and cost, horizontal overflow, page errors and failed requests. These checks do not score the design, accessibility or whether the model got every fact right.
The record. Each published attempt includes its source and input hashes, with model IDs, harness versions and settings where recorded. A completed session means the agent returned a website, not that every part of it works.

Run details

Model ID: Not reported
Started: Not reported
Time: Not reported
Input tokens: Not reported
Output tokens: Not reported
Cost: Not reported
Mobile overflow: None detected
Desktop overflow: None detected
Page errors: 0
Failed requests: 0
Settings: harness: ChatGPT, inputVerification: Archive supplied by Luca. Original prompt and conversation transcript not supplied., integration: Original HTML, CSS and JavaScript unchanged. Hosting files excluded from preview. Sandbox blocks storage, popups and printing., archiveSha256: 9c88eb981cd6f9ea2ac78eb7f5b44ac7b4c3497d8944e8ead1c916964af6d686

Source ↓Run record ↓