The Right AI for The Right Problem

It started with a hackathon: Hangout AI (AI-generated itineraries for Jakarta, Singapore, and Kuala Lumpur) and WIGO AI, and the feeling that LLMs could solve everything. Then came the assessment. For car detection, YOLO + CNN beat GPT-4V at ~$0.001 vs $0.01–0.03 per image and 50 ms vs 3–5 s.

Three case studies follow. Itinerary planning: vector search + OR-Tools gives consistent, optimal plans ~25x cheaper and 16x faster than LLM generation. Price prediction for 100K products: traditional ML (XGBoost / Random Forest) wins on structured data. Content moderation at 10M+ posts a day: custom multi-modal deep learning beats both GPT-4V (cost, latency, privacy) and keyword filters.

The takeaway is a decision framework. Generative → LLM; structured prediction → traditional ML; complex multi-modal patterns → deep learning; optimization under constraints → solver. Then weigh budget, scale, latency, maintenance, and compliance. The best engineering solution is the one that works reliably, scales economically, and solves the real problem.