I build systems that
run without me.

Problem

I run an IELTS institute and needed practice tools for my own students. That turned into a network: practice portals, language hubs, a jobs tracker, a news digest and 275+ utilities — all on one backbone.

What I built

• 359 city guides with census and weather data
• 13 subdomains: IELTS, PTE, six languages, grammar, jobs, news, tools
• One daily pipeline: fetch → validate → build → deploy
• Hub health monitoring, stats and activity feed rebuilt automatically
• Jobs tracker verified and updated twice daily
• Sponsored city-page slots with checkout and auto-expiry

Hard parts

• Programmatic pages without spam: every city page needs real data and a reason to exist. I published a methodology and kept content data-driven.
• One pipeline, many products: shared build logic with per-hub config, so a new hub is configuration, not a new codebase.
• Performance: the 3D network view caps pixel ratio, pauses off-screen and respects reduced motion; content stays in plain HTML for SEO and Core Web Vitals.

Result

• 14 live properties · 359 pages · 13 subdomains
• ~4 months, solo
• Daily automated rebuild with zero manual steps

What I'd do next

Add real usage analytics, remove pages with no demand, and move heavy data jobs to a queue.

Problem

PTE students need repeated practice with feedback. Human evaluation doesn't scale, and generic mock tests don't show progress over time.

What I built

• Question bank across all 22+ task types
• AI scoring for each response
• Progress tracking across attempts
• Practice modes by section

Hard parts

• Scoring consistency: the same answer shouldn't swing wildly. Prompt design, output constraints and fallback rules.
• Cost control: scoring calls are the expensive part — cache and batch where possible.
• Latency UX: show progress state instead of a blank screen.

Result

• Full practice bank live across all sections
• AI scoring and progress tracking in production
• Shipped inside the same four-month window

What I'd do next

Calibrate AI scores against human-rated samples and add weak-area recommendations.

Problem

Most free tool sites are ad farms: slow, tracking-heavy and broken offline. I wanted tools that just work, in a directory that can grow without a backend.

What I built

• 275+ tools: EMI, GST, SIP, percentage, unit converters and more
• Everything processed in the browser
• No signup, no upload, works offline
• Category browsing and search

Hard parts

• Zero-backend architecture: all logic client-side — no server cost, no data leaves the device.
• Consistency at scale: a shared tool template and config so adding tools is fast and uniform.
• Discoverability: category pages and internal linking so 275 pages don't cannibalise each other.

Result

• 275+ tools live
• No signup and no server-side processing
• Built inside the same four-month window

What I'd do next

Per-tool analytics, offline install, and landing pages for high-intent queries.

Problem

"Trending" lists miss early projects and reward old stars. I wanted a signal for emerging work.

What I built

• GitHub API ingestion for 5,200+ repos
• Three-axis scoring model
• Six daily refreshes via GitHub Actions
• Public site and MIT-licensed code

Hard parts

• Rate limits: batch, cache and respect API quotas.
• Scoring that isn't gameable: combine multiple signals instead of raw stars.
• Historical tracking: store snapshots so velocity is real, not estimated.

Result

• 5,200+ repos scored
• Six automated refreshes per day
• MIT licensed and public