TrendLink
LiveMar 2026 – May 2026 · 2 months
TrendLink runs a 4-step LangGraph pipeline on ECS Fargate every evening at 8 PM: it fetches trending IT/tech news via Tavily, uses Claude (via AWS Bedrock) to filter for the most relevant article and generate a polished LinkedIn post, then publishes it automatically — or saves it as a draft in dry-run mode. A React + Vite frontend on Vercel lets users view post history, tweak settings, and trigger manual runs, with all API calls secured through Cognito M2M (client_credentials) JWT authorization on an HTTP API Gateway. The entire infrastructure is managed with Terraform across three modules (api, pipeline, storage), running inside a custom VPC with private subnets and a DynamoDB Gateway Endpoint to keep database traffic off the public internet.
Design Decisions
Chosen ECS Fargate over Lambda for the LangGraph pipeline because multi-step LLM chains exceed Lambda's 15-minute timeout and benefit from persistent in-memory state across graph nodes; EventBridge handles scheduling instead of keeping the container running idle.
Highlights
- ›LangGraph pipeline on ECS Fargate runs daily: Tavily news fetch → Claude relevance filter → Claude post generator → LinkedIn publisher, with early-exit when no relevant article found
- ›Cognito M2M auth (client_credentials flow) secures all API Gateway routes; Vercel serverless functions act as the token-aware proxy layer between browser and AWS
- ›Full IaC via Terraform in 3 modules (api, pipeline, storage) with custom VPC, private subnets, and DynamoDB Gateway Endpoint to keep database traffic off the public internet
Timeline
AI / LLM
Infrastructure
Auth
Other
Tech Stack
Metrics
4 Lambda functions + 3 DynamoDB tables + 3 SSM secrets
4-step LangGraph pipeline (fetch → filter → generate → publish)