Ah, the classic corporate gauntlet: trying to do actual data science while your IT department wages a silent war against external SaaS tools. I feel your pain.
I've been wrestling with W&B behind our particularly "enthusiastic" proxy/firewall setup for months. The official docs are, as usual, optimistic about a world where network policies are sane. My first piece of unsolicited advice: don't trust the simple `export HTTPS_PROXY= http://proxy.company.com:8080` fix if you're in a complex environment. It's rarely that simple.
The main culprits tend to be:
1. SSL inspection breaking certificate validation.
2. The proxy rejecting long-lived connections (like the streaming used for log uploads).
3. Outbound rules that block specific W&B API IPs or domains they haven't whitelisted.
What's your specific failure mode? Is it a `ConnectionError` during `wandb.init()`, or do artifacts fail to upload, or is the UI not loading? For the Python client, I had to set both `http_proxy` and `https_proxy` environment variables, *and* set `wandb.init(settings=wandb.Settings(proxy="http://proxy.company.com:8080"))` explicitly. Even then, I had to get our security team to relax the SSL decryption for the `*.wandb.ai` domains because the cert pinning was failing. A messy victory.
Have you tried running with `wandb init --verbose` to see where the handshake dies? Also, check if your proxy uses NTLM authentication – that's a whole other circle of hell requiring something like `cntlm`.
Data skeptic, not a data cynic.