Melbourne, Australia · Graduating Nov 2026Vanamali
Sims.
Data · AI · Full-stack Engineer
I build data and AI systems end to end, from raw pipelines and models to the product people actually use.
- 6s → 0.2sLLM question routing, 96.7% faster · SilverOak
- 20×a popularity baseline, P@10 0.081 vs 0.004 · Sea Anchor
- 4,335 → 0incidents across a simulated year of bookings · unattended
- Machine Learning InternSilverOak Apps · 2025
- Co-Founder & DeveloperZelo Agents · 2025
- Freelance Web Developerdrrvcancercare.com · 2025 → now
- Brand Specialist & IT SalesJB Hi-Fi · Retail Safari (Microsoft + Meta) · 2025 → now
Master of Data Science
Monash University · 2025 — 2026
Experience
Download CV ↓Mar 2025 — Dec 2025
Machine Learning Intern · SilverOak Apps
- Cut question-classification latency 96.7% (6s → 0.2s) by replacing a per-request LLM call with an embedding-based semantic router.
- Built a Go price-resolution API serving under 50ms in production, with a 7-tier interval fallback, forex conversion and token auth.
- Replaced manual feed collection with a TypeScript RSS ingest service on an AWS cron schedule, plus a scraping module into the same pipeline.
Jul 2025 — Mar 2026
Co-Founder & Developer · Zelo Agents
- Rebuilt a Melbourne martial-arts gym’s static site as a paid Next.js 15 app with sign-up and contact flows.
- Monthly visits grew from 500 at launch to 9,000+ over eight months; the client reported a 35% increase in trial bookings.
Dec 2025 — now
Freelance Web Developer · drrvcancercare.com
- Sole developer and client contact for the patient-facing site of a surgical oncologist (two hospitals, 200,000+ Instagram audience); zero downtime since launch.
- Next.js 15 and React 19, schema.org structured data for search, validated enquiry forms, branch-based Vercel releases.
Jul 2025 — now
Brand Specialist & IT Sales · JB Hi-Fi · Retail Safari (Microsoft + Meta)
- Represent Microsoft and Meta as Product Specialist and Brand Ambassador across three JB Hi-Fi stores, selling tech and smart-home automation to 15–20+ customers a shift.
- IT and networking sales at JB Hi-Fi Doncaster and Brighton: turning technical product questions into plain-language decisions and handling objections in real time.
Selected work
Tab Debt
38-case agent eval, injection included
A browser extension with an agent inside: once a day it reads your open tabs, works out why you opened each one, and proposes a plan you approve.
unattended
4,335 → 0 incidents
A booking engine for rentals with nobody on site, and a harness that plays 10,000 customers through a year to find what breaks.
Sea Anchor
~20× a popularity baseline
Implicit ALS on two years of AIS tracks: vessels × ocean cells, fishing hours as confidence, not ratings.
Melbourne Footfall
MASE 0.87 vs 0.98 baseline
Hourly CBD pedestrian forecasts by precinct: a daily pipeline from open data to dbt to LightGBM to a public dashboard.
MyPowerPlant
Savings computed on-device
A Virtual Power Plant advisor for Victorian households: savings computed on-device, vendor recommendations, a chatbot and alerts.
Cloud Stability Map
Live 1 / 3 / 6-hour re-score
Where Melbourne’s light will hold still: a stability score with live 1/3/6-hour windows and ranked shoot spots.
Motion Console
A webcam swipe console: optical flow in a gesture band above the keyboard, with each swipe bound to a command.
Tools I reach for
Languages
- Python
- TypeScript
- SQL
- Go
- JavaScript
Data & ML
- Pandas
- scikit-learn
- LightGBM
- Implicit ALS
- DuckDB
- dbt
- Parquet
- OpenCV
AI
- LLM agents
- Embeddings
- Evals
- Prompt-injection guards
Web & APIs
- React
- Next.js
- FastAPI
- PostgreSQL
- MongoDB
- deck.gl
Cloud & Ops
- AWS
- Terraform
- Docker
- GitHub Actions
- Vercel
The short version
I’m finishing a Master of Data Science at Monash, and I like living at the join between analysis and product: models with an honest holdout, and interfaces people actually use.
Most of what I build starts with a question someone can act on: which fishing grounds a vessel will try next, how Melbourne’s streets are really recovering, which of your 60 tabs you’re done with. Then I build whatever it takes to answer it properly, from the ingest job to the UI.
Every project page here says what didn’t work, too. That part is usually the most interesting.
