Our fourth weekly catchup. We organize an agenda the day beforehand and consolidate our findings, notable research progress, possibly new tech, insight or technique everyone should know.

Agenda:

• Broadway in elixir to handle million messages by @Hieu
• Using WebAuthn API to implement passkeys by @Vincent
• Labeling image using openAI vision and embedding flow by @Lap

Output:

• Understand the research progress and passions of the labs team
• Understand any shortcomings or things to improve for the next weekly catchup

Notes from the catchup

weekly meeting #4 - 27/12/2023:

⚖️ Using Broadway in Elixir to Handle millions of messages

• Broadway is a concurrent, multi-stage tool for building data ingestion and data processing pipelines in Elixir.
• @Hieu mentions that we have a backlog issue for requests through NgheNhan’s websocket, and to overcome this limit, we can introduce Broadway to help manage backpressure for requests.
• @Hieu mentions that we can use it to manage process groups to manage pipelines such as for getting market prices.
• (@Hieu) Managing and scaling across multiple nodes will require a consensus algorithm such as Raft to handle configurations between nodes and message handling

🔒 Using WebAuthn API to Implement Passkeys

• A joint effort between @Vincent and @An Tran
• (@Vincent) We can use a WebAuthn, a growing web standard, to help simplify authentication for users: https://webauthn.wtf/
• (@Vincent) Passkeys are generated on the machine and can be stored locally, or it can be stored in another device
• (@Vincent) By default, there isn’t an open external service to help store passkeys for the user, apart from third-party sources such as 1Password and iCloud
• (@Vincent) Implementing WebAuthn is as simple as the processes we use for authentication Metamask on Mochi UI

🏷️ Labeling image using GPT4 Vision and Embedding flows

• (@Lap) In one of our projects, iFramely to help us get the initial metadata and we then use GPT4 vision to help label images
• The labels are then used in generating embeddings for related text and the image itself, regenerate inputs for classification, and used to create a suggestion record with weights.
  • Cohere is used to do text embeddings where as openai/clip-vit-base-patch32 is used for image embeddings, both of which help store multi-dimensional vectors
• (@Thanh) There does seem to be a 3-4 round trip process before finally getting enough data to create suggestion scores for the user