deploy AI server to AWS

overview

building the ai speech server was only half the job so i pushed the fastapi service onto AWS EC2 and wired a github based delivery pipeline to keep deployments hands off

---

background

• requests
  • players talk to the golem npc through voice inside the local build so the fastapi speech api must live on a remote endpoint
  • the previous local setup has to stay reachable from anywhere
• priorities
  1. host the fastapi service on the remote node
  2. prepare the AWS EC2 environment
  3. add a github deploy workflow
• lower priority
  • harden secrets with .env files kept outside docker images
  • stop exposing the EC2 ssh port to every ip

purpose

make the speech to text server land on the AWS instance automatically with zero manual patching

environment setup

• dependency stack managed with uv and pyproject.toml
• server docs endpoint: http://44.210.134.73:8000/docs
• websocket endpoint: ws://44.210.134.73:8000/api/v1/ws/speech/v1

---

implementation

ec2 instance

Image hidden for security reasons

design

• added a github action near the repo root

.github/
└── workflows/
    └── deploy.yml

• deployment workflow reference
  https://github.com/ob1hnk/Triolingo/blob/main/.github/workflows/deploy.yml

result

• accessing the instance through the pem key shows the container running

ubuntu@ip-*:~$ docker ps
CONTAINER ID   IMAGE                                              COMMAND               CREATED        STATUS        PORTS                               NAMES
9b20bf338df7   ghcr.io/ob1hnk/triolingo/gogo-golem-ai:main-fec67ba   "uv run uvicorn..."   14 hours ago   Up 14 hours   0.0.0.0:8000->8000/tcp [::]:8000->8000/tcp   gogo-golem-ai-server

• confirmed the websocket api responds end to end

pr

https://github.com/ob1hnk/Triolingo/pull/2

---

troubleshooting

• issue 1
  docker refused any uppercase letters in the image tag even though IMG_NAME was lowercase the whole time in the workflow reruns the old commit kept executing because rerun failed job sticks to its original sha this made the fix look ineffective check the screenshot log for proof
  

• issue 2
  after editing cached image names the pushed tag and the pulled tag no longer matched so ec2 tried to fetch an image that did not exist i standardized everything to main-{short_sha} with seven characters to kill the mismatch
  reference commit: https://github.com/ob1hnk/Triolingo/commit/f3edcc47965ff26b92be86c143f645b9e842dd4c

reference

https://github.com/orgs/community/discussions/27083