--- title: "Recce Cloud: Hosted Data Review Agent" description: "Case study: building the Recce Cloud platform on AWS ECS Fargate. CDK-driven secrets pipeline, HMAC-verified Slack receiver with idempotent event handling, ephemeral container lifecycle, and coordinated major-version migrations across the stack." doc_version: 1 last_updated: 2026-05-24 canonical: https://variable.team/projects/recce-cloud --- # Recce Cloud - **Company:** [Recce](https://reccehq.com) - **Industry:** Data Engineering / Analytics Engineering - **Dates:** Apr 2025 to Present - **Project link:** - **Stack:** AWS, CDK, Python, FastAPI, SQLAlchemy, ECS Fargate, RDS, Next.js, TypeScript, Slack API ### Wiring secrets from GitHub through CDK to running containers The Ask Recce Slack bot needed an Anthropic API key, a Slack signing secret, and a GitHub dispatch token, all delivered through a chain that staging and prod could differ on. I wired the path end to end: GitHub Actions secrets, CDK Python pipeline stages, ECS task definitions, container environment variables, and the FastAPI runtime config. The pattern fails fast when a secret is missing rather than producing the silent auth failures that had bitten earlier integrations. ### A Slack receiver that doesn't double-process events Slack retries events when it doesn't get an ack in time. The Recce Slack receiver handles HMAC signature verification, strips bot mentions before parsing, and deduplicates on the Slack event ID so re-deliveries don't kick off a second analysis run. The receiver chains into a GitHub Actions dispatch, passing the parsed payload through with type-safe Pydantic models. Tests cover the missing-secret and malformed-payload cases. ### Ephemeral container orchestration for preview, share, and task instances Recce Cloud spins up containers on demand. PR preview environments, shareable analysis sessions, and one-off agent tasks all run the same way. The lifecycle abstraction uses Docker locally for dev and ECS Fargate in production, exposing one interface so application code doesn't branch on environment. Health checks, port allocation, and lifecycle events all sit behind the abstraction. ### Coordinating major-version migrations across the stack The Cloud stack absorbed several breaking upgrades over the year: pnpm 10 to 11, mypy 1 to 2, Starlette 0.x to 1.0. I coordinated those across 7 GitHub workflows and 2 package manifests, validated compatibility against frozen lock files, and rolled the changes out in chained PRs so any regressions stayed contained to one stage at a time. ## Sitemap [Full site index](/sitemap.md)