Portfolio & Case Studies

Most CS learners never ship anything — they finish tutorials but have no portfolio, no proof they can build. Levelio reframes the unit of progress: instead of lessons completed, it's projects shipped. Each level unlocks the next real-world build, with AI guidance adapting to your current skill level.

• ▹Analysed drop-off patterns across 5 competitor platforms — identified the exact moment users disengage (post-tutorial, pre-first-project)
• ▹Designed AI guidance layer: context-aware hints, adaptive difficulty, and portfolio-progress tracking to keep learners in flow
• ▹Wireframed core user flows in Figma: onboarding, project unlock, and AI hint panel — optimised for mobile-first engagement

Manual expense logging fails because it's too slow — so people stop after a week and have no idea where their money went. Built a Python app that reads receipts via OCR, extracts vendor, date, and amount automatically, and visualises spending trends in a live Streamlit dashboard. Goal: zero-effort financial awareness.

• ▹OCR pipeline reduced data entry from ~8 manual fields to 1 action (photo upload) — removing the main friction point entirely
• ▹Dashboard surfaces category breakdowns and monthly trend lines that were previously buried in raw data — built for insight, not data dump
• ▹Designed for Southeast Asian receipt formats — tested against Khmer and English receipts with varied layouts

Businesses making inventory decisions on gut feel either overstock or run dry. Built an XGBoost model with 43 hand-engineered features, benchmarked against ARIMA and LSTM to find the right accuracy/complexity tradeoff. The goal was a model that a non-technical manager could actually use.

• ▹Engineered 43 features — ~40% accuracy gain over baseline; 11.6% MAPE, 0.856 R²
• ▹Benchmarked XGBoost vs. ARIMA vs. LSTM; XGBoost won on accuracy and interpretability for this dataset size
• ▹Dashboard makes forecasts readable without touching the model — designed for ops managers, not data scientists