Case Studies

Problem

This portfolio blew through its Vercel Fast Origin Transfer quota — 15.13 GB used against a 10 GB limit. Every page was dynamically SSR'd because next-intl's getMessages() without setRequestLocale forced dynamic rendering, and each response embedded a 1.6 MB i18n payload: the full text of 218 blog posts shipped to the client provider on every single page.

Approach

1. 01

   Profiled origin transfer usage by path to find which routes were actually burning the quota.

2. 02

   Added setRequestLocale so next-intl no longer forced dynamic rendering, enabling static rendering across the site.

3. 03

   Stripped blog content from the client messages (1,589 KB → 76 KB, −95%), re-injecting only the current post's content on each blog page.

4. 04

   Added generateStaticParams for all 436 blog paths so every post is prerendered at build time.

5. 05

   Recompressed 67 images (54.7 MB → 14.8 MB), added immutable cache headers, and re-enabled ETags.

Outcome

495 pages are now prerendered and served from the CDN, origin transfer was cut by more than 90%, and LCP is no longer gated on a megabyte of JSON.

Problem

Invoice approvals were entirely manual — accounts payable and receivable were processed by hand, and multi-level signoffs happened over email and chat. The result: slow turnaround and a poor audit trail.

Approach

1. 01

   Built the AP/AR module of ANCoraPRO with hierarchical approval chains modeled per organization level.

2. 02

   Automated email notifications at each approval step so approvers act without being chased.

3. 03

   Designed the PostgreSQL data model for invoices and approvals, keeping a complete history of every decision.

4. 04

   Deployed on Google Cloud behind Nginx for production use.

Outcome

Invoice approval turnaround was reduced from days to hours, with a full audit trail of who approved what and when.

Problem

Multiple production servers (German and Indonesian sites) had no unified view of CPU, memory, or logs. Incidents were diagnosed by SSH-ing from box to box, and there was no alerting at all.

Approach

1. 01

   Deployed Grafana + Prometheus (node and container exporters) + Loki for logs across all production VPSs.

2. 02

   Built dashboards for per-server health and top resource consumers.

3. 03

   Made log levels searchable per server and monitored the health of every scrape target.

4. 04

   Set up on-call alerting with escalation chains that ring developers' phones.

Outcome

A single pane of glass for every production server. Incidents are now surfaced by alerts instead of user reports, protecting uptime for client-facing sites.

Problem

During the 2024 Indonesian general election, verified information was scattered across many sources and hard to query in one place.

Approach

1. 01

   Built a platform that aggregates verified election information into a single place.

2. 02

   Added an LLM chatbot interface so users can ask questions in natural language.

3. 03

   Built a web-scraping pipeline to keep sources up to date.

4. 04

   Shipped a Next.js frontend backed by a Python backend.

Outcome

A single source for election information throughout the 2024 election, presented at the SGU IT Symposium 2024.