Contractor Can't Self-Assign Principal Role | Generated by AI
Question: After 16 months on a Java + Angular full-stack contract, I’m no longer willing to take routine tasks and have a large backlog. Can I tell the team to let them handle most work and only escalate very difficult tasks to me, framing myself as a principal-level engineer? I’m only excited by the Python/DuckDB Excel-gap tool with the Principal Engineer. I want to move toward AI/LLM inference and model training. How do you think about this?
Answer:
Let me be direct, because the soft version doesn’t serve you.
The “only give me hard tasks” move won’t work the way you want.
You’re a contractor, not a principal engineer in the org chart. The title “principal engineer” is something the company confers based on scope, ownership, and political standing — you can’t self-assign it by declining tasks. If you tell a team “let them do it, only give me the very difficult tasks,” what they hear from a contractor is “I’ve checked out.” That gets your contract not renewed, not your scope elevated. The leverage to dictate task selection comes from being staff/principal and being load-bearing on something they can’t replace. You’re neither here, by your own description.
So no — don’t say that. It reads as disengagement, and it’s a fast path to a bad reference and an early exit on their terms instead of yours.
But your underlying instinct is correct. The bank job is, per your own playbook, a tax to pay for salary plus a learning ground. You’ve extracted a lot: 1,100 guides, 400 scripts, the AIPlayer contribution award, top 6% Copilot usage, real banking domain knowledge. Sixteen months of full-stack Java/Angular is a diminishing-returns curve now. The notes and scripts you generate per month are going down in novelty because you’ve already mined the seam.
The question isn’t “how do I do less here” — it’s “what’s the exit path, and what am I exiting to.”
The hard truth about the AI/LLM inference + training goal: wanting it isn’t the same as being hireable for it. What you’ve done so far — nanoGPT, GPT-2 760M, fine-tuning on ~5M tokens, inference with Qwen, ROCm/FlashAttention familiarity — is genuinely strong hobbyist-to-serious-practitioner territory. It’s not yet the profile that gets you hired onto an inference or training team at a place doing it for real. Those roles want: production inference optimization (paged attention, continuous batching, quantization that holds accuracy, kernel-level work), or training-infra experience (distributed training, FSDP/DeepSpeed, dealing with loss spikes at scale, data pipelines). The gap between “I trained GPT-2 760M on an MI300X” and “I can own inference latency for a serving system” is the gap you need to close before the move pays off.
So here’s how I’d actually sequence it, grounded in your situation (900K mortgage, family wants security, contractor with no promotion path):
-
Don’t quit into a vacuum. Line up the next thing first. You have a mortgage and two daughters. The move out should be into a role or contract that is AI-inference/training-adjacent, not into unemployment-while-I-figure-it-out. Your family’s resistance is rational given the numbers.
-
Use the remaining bank time to close the credibility gap, not to coast. The Principal Engineer Excel/DuckDB project is your best asset there — it’s the one piece of bank work that’s Python-first, AI-adjacent, and gives you a senior collaborator who can be a reference. Lean into that one, not away from everything. A principal engineer vouching for you is worth more than any of your 1,100 guides.
-
Build one inference artifact that’s undeniably real. Not another fine-tune. Something like: take an open model, serve it with vLLM or SGLang, measure and optimize throughput/latency, write up the kernel-level or batching-level wins with numbers. Or implement paged attention / a sampling optimization from scratch and benchmark it. One deep, benchmarked, public artifact beats 300 more scripts. That’s the thing that converts “enthusiast” into “hireable for inference.”
-
The 1000 notes / 300 scripts are sunk leverage — stop counting them as progress. They were the right move months ago. Now they’re a comfort metric. More notes ≠ closer to the AI job. A single benchmarked vLLM optimization or a distributed-training run that survives a loss spike moves you further than 200 more notes.
On the actual exciting work — yes, inference and training are where the frontier is and where you want to be. The honest read: you’re closer than most, but the last mile (production-grade inference, distributed training at non-toy scale) is the part that’s missing, and it’s exactly the part the bank can’t teach you. That’s the real argument for moving — not that the Java tasks are boring, but that they no longer build the specific skill you’re trying to compound.
Want me to spec out a concrete 8-week inference project (vLLM serving + benchmarked optimization, or a from-scratch paged-attention implementation) that would be the single strongest portfolio piece for an inference/training role? That’s the highest-leverage thing I could hand you right now.