Speaker #1 - Richard Lee

multiple agent system executive agent controlling many child agents child agent must be very silo and purpose built

key challenge —> hallucination solution —> multiple layers of grounding

Dont overengineer—> monolith might be better Dont jump straight into micro-services you can always transit to scale later speed is still important

if high risk/ destructive ops, always go for deterministic approach (code)

never roll your own auth! airlock your auth via containerising

observability to consider grafana for observability

Speaker #2 - Alec Teo on PixelPuggles

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

supabase auth vs clerk for auth to create an auth after 3 plays for lower barrier of entry

natively integrated with supabase
tradeoff is that supabase auth message is not as great
clerk in terms of uptime is not as good as supabase

data schema is still very important in system design, need to review it yourself

to allow agent interaction with your app to prep for an agentic future

alex manually pixalate the photo and batch upload the photos can consider offloading pixalation on the user’s device to better optimise

in terms of image gen costing consider imagen 4 (6x cheaper) over nano banana2

try isitagentready.com to make it more agent friendly

To consider automating part of the image generation eval

LLM as a judge to check (must be a different model from image gen model)
have access to a scientific database for best accuracy

Speaker #3 - Dalton

https://www.linkedin.com/in/dalton-prescott-ng/

https://github.com/evokoa/pggraph

pgGraph_slides.pdf

graph db is very expensive cost is 34x cheaper than neo4j

Database hydration (also known as object hydration) is the programming process of converting raw database query results (like a row or an array) into fully functioning, in-memory programming objects

Big usecase would be fraud detection

RAG vs Graph RAG —> good for search Graph —> find out relations between entities

More photos here: - (to be retrieved from telegram later on)