High-Scale Engagement and Multimodal Labeling Platform

Problem

The company needed two things that look unrelated until you share infrastructure: a social engagement product that spanned Web2 and Web3 identity, and a labeling pipeline that could ingest multimodal media at the volume expected by large AI and media clients. Neither problem is interesting as a demo. Both become interesting at concurrency, identity, and worker throughput.

Constraints

The engagement product had to feel like a social graph, not a ledger explorer. The labeling path could not be bound to a single browser session if it was going to hold client-scale volume. Services had to deploy and scale independently. The two products could share identity and infra patterns without sharing product surfaces.

Architecture

A graph-backed social model in Neo4j, exposed through a GraphQL API. Kubernetes services owned the write and read paths so engagement and labeling could scale on their own axes. Electron desktop workers connected to a central indexing and model system, which meant labeling throughput was not trapped in a tab. The engagement product and the labeling product shared identity and cluster patterns; they did not share UI.

Role

I worked across the GraphQL surface, the graph model, the Kubernetes services, and the worker connection into the central indexing system. The job was to keep the social product coherent while the labeling path took on volume that a browser-only pipeline would not survive.

Outcomes

• 400,000 active users on the engagement product.
• Labeling pipeline operated at the throughput expected by clients at Adobe / Google / OpenAI scale. That is a statement about operating volume, not a personal endorsement by those companies.
• Desktop workers detached labeling from the browser so throughput could grow with the worker pool.

Stack