My day job. Around 2,500 Australian school principals take a wellbeing survey each year. This pipeline turns every submission into a personal PDF report, and flags anyone whose answers suggest they need support. I work on the code that generates them.

How it works

REDCap survey submitted
  -> score the responses
  -> compare to the person's own past
     years + national benchmarks
  -> render a PDF with charts
  -> email it out
  -> in parallel: risk check, escalate
     to the research team if triggered

What I've shipped

• Rebuilt the report renderer for the 2026 design, running beside the old one so nothing breaks mid-study
• Fixed a silent data bug where charts pulled the wrong file depending on which directory the job started in
• Wrote the documentation site (MkDocs), the first written docs the pipeline has had
• Audited the scoring logic and filed 18 defects, 4 of them affecting what participants actually read

The hard part: those ~5,000 PDFs render unattended and nobody proofreads them before they send. A quiet scoring bug reaches a real person.

Runs under HREC ethics approval, so: data minimisation, no participant data into third-party tools, every change peer reviewed. Private repo, no links.

Built and run it on my own. The public site for UTS Bangladeshi Society, serving a community of ~500. MIT-licensed.

The committee changes every year and none of them write code, so the site is built to be edited by people who aren't developers: content comes out of a CMS, and every change gets a preview deploy before it goes live.

- Next.js 16 App Router + React 19
- Prisma + PostgreSQL on Supabase
- The Graamys: awards voting across
  9 categories, nomination to result
- CMS-driven pages + event galleries
- Vercel CI/CD, preview per PR

Any AI tool will write you a research report. This one won't publish a sentence it can't trace back to a real source.

Give it a topic. It searches, saves an exact copy of every page it reads, then writes the report only from what those pages actually said. Each claim gets checked back against the archive. Claims that don't hold up are withheld and listed with the reason, not quietly left in.

search -> archive (WARC) -> extract
   -> draft from facts only
   -> check every claim against
      the archive
   -> ship what holds, name what doesn't

Any run replays offline from its own archive, so you can re-check a report months later. Also runs a separate numeric check, because a model that gets a number wrong usually still sounds confident.

Built on crawl4ai. Published to PyPI; source private while the archive layer settles.

An assistant that knows who you are, so your tools stop asking.

I had six automation scripts job-hunting for me. They applied to zero jobs for six days. Nothing crashed: they filled in 13 of 20 form fields, hit "do you need visa sponsorship?", and binned the application. The answer was already in a config file. Nothing thought to look, and the alert about it sat unread in a text file for 143 hours.

aibo brief    what happened, what is stuck
aibo doctor   health measured by outcome,
              never by "the job ran"
aibo asks     the queue only a human can
              answer  (--push to your phone)

Every fact it stores records where it learned it. It answers on your behalf only when it can point at where you already said that. Everything else becomes a question for you.

ENGG2000/3000 group capstone. A working model drawbridge that raises and lowers itself, built to fail safely.

Two microcontrollers, split so neither can strand the bridge. They ping each other constantly; if either stops answering, the bridge stops moving.

- ESP32 <-> Arduino over serial
- Heartbeat between the two boards
- Emergency stop that cuts drive
  regardless of what software wants
- Manual override + automatic mode
- Weight and position sensing

Cold outreach, automated. Type a company name. It finds the right people, drafts a personalised email to each, and queues them for you to approve before anything sends.

- Contact lookup across sources
- Claude API writes the drafts
- Job queue (BullMQ + Redis) so a
  slow lookup can't block the app
- Stripe billing, Supabase Auth

Frozen as a portfolio piece (April 2026). Clone and run npm run demo, no API keys needed.

Motion Control (2024). A motor-driven carriage that runs a set sequence and stops itself before it hits anything. Most of the work was in the sensors: raw readings are noisy, so the code has to tell a real obstacle from a bad reading.

RFID Inventory Tracker. Team project, I was Scrum Master. Scanned tags update stock in real time and sync automatically. I ran the sprints and the sessions with users to work out what they needed.