AWS Builder Center
NoMeshOps

NoMeshOps

A self-healing deployment agent that doesn't just report why a deploy failed — it diagnoses the failure, fixes it, and verifies the fix for real, using a rules table, a DynamoDB knowledge base, and Amazon Bedrock only as a last resort.

Gen ai engineer

NoMeshOps: a deployment agent that fixes its own failures

Every developer knows this moment: you deploy to a fresh machine, and it fails on something that has nothing to do with your code. A missing system library. A package that needs to compile from source and has no compiler. An app that installs cleanly and then crashes the instant it starts. You SSH in, read the log, run one shell command, redeploy, and repeat — sometimes three or four times before it works.
NoMeshOps is an agent that does that loop itself. You give it a target machine and a Python repo. It deploys the project, and if the deploy fails, it classifies the failure, tries a fix, and verifies the fix actually worked — install succeeded, every package actually imports, the app process is running, and the health endpoint answers from the process it just started, not a stale one still holding the port — before it trusts the fix and moves on.
It was built for AWS's WeMakeDevs First Commit hackathon, Ship It track.

The core idea: cheap before expensive

The interesting design decision isn't "call an LLM when something breaks." It's the order NoMeshOps tries things in, and the fact that it verifies every single one before believing it:
  1. A deterministic rules table. Runs in well under a millisecond, touches no network. Most deploy failures are things that have been seen a thousand times before — a missing  git  binary, a source build that needs  libpq-dev , a Python venv module that isn't installed on a minimal image. These get fixed instantly, for free.
  1. A DynamoDB knowledge base, keyed by an error signature — a hash of the error type, the package, the OS, the architecture, and the Python version. The same error text on Python 3.9 and 3.12 is treated as a different problem on purpose, because it often needs a different fix. A family-index GSI lets a fix discovered on one platform get a second chance on another, but only through the same verification gate — nothing is trusted just because it worked somewhere else once.
  1. Amazon Bedrock (Claude, Converse API) — only when both of the above miss. One JSON object per call, under a strict contract, checked against a destructive-command deny-list before anything reaches the target. Capped at one model call per run. This is the expensive, slow, last-resort path — which is exactly how it should be used.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article