AI Infra Career Survival Plan | Generated by AI

Home 2026.07

Question: Given my situation (AI infra transition, no formal degree, 900k CNY mortgage, leaving bank contractor job tomorrow), how should I survive and plan my career?

Answer:

First: do not think of yourself as a fresh AI engineer. Your position is unusual.

You have:

11 years engineering experience
backend/distributed systems experience
cloud/Linux/Kubernetes
real LLM training experience:
- GPT-2 124M from scratch
- nanoGPT/nanochat experiments
- vLLM exploration
- GPU debugging
- CUDA/ROCm exposure
shipped products before

The problem is not capability. The problem is market signaling.

A degree is a filter for HR. It is not the core requirement for AI infra.

But with 900k mortgage, you cannot make a romantic “quit everything and research AI” move. You need a two-track strategy.

1. Next 3 months: survival mode

Your priority:

Cash flow > perfect AI job

Do not burn runway.

Your target:

Maintain income: 20k-40k RMB/month
Move toward AI infra every month

Possible paths:

Path A: AI Infra Engineer (best)

Roles:

LLM Platform Engineer
Inference Engineer
ML Infrastructure Engineer
AI Backend Engineer
Agent Infrastructure Engineer

Your existing backend background maps well.

The pitch:

“Backend engineer transitioning into LLM infra. Built and trained GPT models, deployed inference systems, optimized GPU workloads.”

Not:

“I have no degree and want to learn AI.”

Path B: AI consulting / outsourcing

You already have an advantage.

Many companies need:

RAG systems
private LLM deployment
agent workflows
internal AI tools

A small company may pay:

30k-80k RMB/project

Your engineering background is valuable here.

Path C: Keep a bridge job

A mistake would be:

bank contractor
     |
     v
unemployed
     |
     v
study AI 6 months
     |
     v
job hunting

Mortgage makes this dangerous.

Better:

bank contractor
      |
      +---- freelance AI projects
      |
      +---- AI infra portfolio
      |
      v
AI infra job

2. Your AI infra roadmap

Do not learn more models.

You already know enough theory.

Now build production evidence.

Month 1

Build:

Project 1: Mini vLLM clone

Implement:

prompt
 |
tokenizer
 |
KV cache
 |
continuous batching
 |
scheduler
 |
CUDA kernel
 |
generation

Even a toy version.

Repository:

github.com/lzwjava/my-vllm

README:

Implemented:
- PagedAttention
- KV cache management
- batching scheduler
- tensor parallel inference

This signals AI infra.

Month 2

Build:

Production LLM serving stack

Example:

Client
 |
FastAPI
 |
Ray
 |
vLLM
 |
NVIDIA GPU
 |
Prometheus
 |
Grafana

Add:

autoscaling
monitoring
latency benchmark

Measure:

tokens/sec
TTFT
GPU utilization
memory usage

Month 3

Contribute upstream.

Targets:

vLLM
llama.cpp
PyTorch
Triton

Even small PRs.

A single merged PR can replace a degree signal.

3. Degree problem

The reality:

Large companies:

Google
Meta
OpenAI
Anthropic
NVIDIA

often use degree filters.

But AI infra ecosystem is different.

Many teams care about:

Can you make GPU cheaper?
Can you increase throughput?
Can you debug CUDA?
Can you deploy models?

Your advantage:

Most people who have degrees cannot do what you already did.

A person with CS PhD but no production experience may not know:

nvidia-smi failure
CUDA mismatch
flash-attn build
distributed training
Linux kernel issues

You do.

4. Financial plan with 900k mortgage

Assume:

Mortgage:

900k RMB

Don’t optimize for maximum salary first.

Optimize for:

survival probability

Your hierarchy:

Level 1

Stable income:

20k+/month

↓

Level 2

AI transition

↓

Level 3

30k-50k/month AI infra

↓

Level 4

Consulting/startup

Avoid:

buying more GPU hardware
spending 6 months only reading papers
chasing PhD-level theory
quitting without income

You already have enough compute.

Your RTX 4070 is enough for portfolio work.

5. Your story matters

Your background is actually a strong founder/engineer story:

self taught programmer
 |
mobile/backend engineer
 |
built products
 |
learned deep learning
 |
trained GPT models locally
 |
building AI infra

This is closer to:

Karpathy style
independent researcher-engineer

than traditional graduate student.

Use it.

6. My recommended next 12 months

Aug-Sep 2026

Get income bridge
+
build vLLM-like project


Oct-Dec 2026

Apply:
- AI infra engineer
- LLM platform engineer
- AI backend engineer

Contribute OSS


2027

Choose:

A:
Senior AI infra engineer

or

B:
AI consulting company

or

C:
AI startup

The key decision tomorrow is not “should I leave the bank job?”

The key question is:

After leaving, how many months of runway do you have?

Calculate:

cash savings / monthly family expense

If runway < 6 months:

Get income immediately while transitioning.

If runway > 12 months:

You can aggressively build AI infra portfolio.

Given your mortgage, I would personally optimize for income + AI transition simultaneously, not pure exploration. You are already past the beginner stage; the next step is turning your AI ability into market proof.

Back Donate