Roast #3 · April 17, 2026 · Singapore

• Image-heavy PDFs: extract into a.md format with Docling.
• Text-based PDFs: extract into a.md with MarkItDown (Microsoft).
• Python alternative: pdfplumber.

Use deterministic code for numeric analysis — less flexibility, but you keep a clean audit trail. And don't do a real-time API call if you can help it: pre-compute and cache instead.

Pick a different model per task: use gpt-4o mini for categorization and simpler work, and a more conversational model to handle negotiation and distill negotiation style. Consider shadcn for ease of UI design.

The audience showed two paths: adopt DuckDB for analytical workloads, or handle idempotent writes in Postgres with upsert / on conflict patterns.

The strength of this approach is that it mimics functional programming — data flows through pure steps instead of mutating in place. Which raises the natural question: why not use a functional language for the whole stack?

For the API layer, build with FastAPI instead of rolling your own with Flask — notably more maintainable.

A key speedup: the disk write happens off the critical path, so there's no user-visible latency. This only works because everything is immutable — once written, nothing needs to overwrite or coordinate on the same bytes.

• Remotion — build and render programmatic videos using React code.
• Voicebox — a local tool that clones your own voice for personalized audio.
• FFmpeg — the versatile framework to decode, encode, and process video and audio files.
• Google Gemma — lightweight, open models capable of video summarization tasks.