Roast #4 · May 7, 2026 · Singapore

• API — the first-class route.
• Scraping — when the API doesn't give you what you need.
• 3rd party analytics — as a complement or fallback.

E → Extract (bronze) · T → Transformation (silver) · L → Load (gold). Databricks is the cloud-based platform that runs the data analytics and AI workloads.

To overcome time-out issues, consider retries with backoff — especially for LLM calls and APIs. This prevents a huge bill shock when a provider starts failing in burst.

The One Big Table approach beats a relationship database (e.g. Neo4j) for query speed and simplifies data consumption — fewer joins, one place to ask questions of.

If you have a single source of truth, RAG is preferred over a raw LLM — it increases reliability. And if you need the LLM to do a web search, use a dedicated web search tool rather than letting the model guess.

Consider an n8n dynamic model selector to choose the best model per use case. n8n reduces LLM usage overall and can switch between using an LLM web search and RAG depending on the query.

Consider Postgres for the RAG store. For cost efficiency, consider other LLMs for this use case — like GPT-5.4 mini or other open-weight models.

Other tools raised: the eval tool in the Claude platform, and Kapso for WhatsApp business workflows.

ASR (automatic speech recognition) turns the interview questions into text, and Whisper does the text-to-speech for the answers. The result: employers talk to a resume like it's a candidate.

For a budget option, Oracle Cloud's free tier lasts forever: 24GB RAM, 4 CPU, and 10TB egress — plenty for a side project with a video-heavy workload.

Always consider guardrails for AI applications, and the simplest way to build them is via a persona.md — encode who the AI is and what it will not do right in the prompt, checked before any answer ships.