Speaker #1 - Fliou on TransPrice

https://www.linkedin.com/in/abdelrahman-fliou/

TransPrice_RoastMyTechStack.pdf

Ways to Parse PDFs to Markdown files

1. Extract Image PDF into a.md format https://www.docling.ai/

2. Extract text PDF into a.md format https://github.com/microsoft/markitdown

3. Python alternative: https://github.com/jsvine/pdfplumber

Use deterministic code for numeric analysis (less flexibility but ensure audit trail)

Dont do a real-time API call if you can help it

Consider shadcn for ease of UI design https://github.com/shadcn-ui/ui

Selective of model for different tasks: → use gpt4o mini for categorization and simpler work → use a more conversational model to handle negoatiation and distill negotiation style

Speaker #2 - Gaurav

https://www.linkedin.com/in/gauravmanek/

https://www.gauravmanek.com/lectures/2026/roast-my-tech-stack/

VisigothAI Data Storage-compressed.pdf

• IF you want to switch to a relational DB, you can either:
  • Use DuckDB as a relational DB engine
    • https://github.com/duckdb/duckdb
  • Design the rows with upserts, and cache on version numbers (instead of references)
    • https://stackoverflow.com/questions/17267417/how-to-upsert-merge-insert-on-duplicate-update-in-postgresql#:~:text=PostgreSQL doesn't have any,unique keys cause unique violations.
• The strength of the approach is that is mimics functional programming - so why not use a functional language?
• Build the API with FastAPI instead of rolling your own with Flask. It makes it a lot more maintainable. https://github.com/fastapi/fastapi
• A key speedup is that the disk write happens off the critical path, so no user-visible latency! Only possible because everything is immutable.

Speaker #3 - Arul

https://www.linkedin.com/in/arulkumaran/

https://arul.sg/