We’re building deeply integrated LLMs into a real product used daily by restoration companies running thousands of jobs. This is not a “prompt engineer” role. You’ll design, train, and ship domain-specific language models that automate real workflows and move real revenue.
Responsibilities
You will:
•
Own end‑to‑end LLM systems: architecture, training, evals, and iteration
•
Fine‑tune and extend existing models (LoRA, instruction tuning, RLHF)
•
Build and maintain data pipelines from product databases, documents, APIs, and logs
•
Ship reliable, monitored, production models with clear guardrails
•
Collaborate closely with product and engineering to turn messy real‑world problems into working systems
•
Build and coordinate the AI engineering team
•
Use Claude Code as a core tool for development, refactors, tests, and experiments
Requirements
This is for you if:
•
“How does this actually work under the hood?” is your default question
•
You’re fine sitting with a hard problem for days and reading papers on weekends to figure it out
•
If there’s something interesting to learn or solve, it doesn’t matter if it’s Saturday or 1 a.m., you’re in
•
You build side projects nobody asked for and write cleaner code than anyone requires
•
You’re quietly competitive, self‑taught in at least one major skill, and think in systems
•
You’re slightly allergic to meetings without a clear purpose or owner
•
5+ years of real world experience in ML / AI engineering
•
Proven experience training or substantially contributing to training LLMs (not just calling APIs)
•
Deep understanding of transformers, attention, and training dynamics
•
Strong Python plus PyTorch or JAX
•
Experience with large‑scale data pipelines and experiment tracking
•
Hands‑on fine‑tuning (LoRA, instruction / SFT, RLHF or similar)
•
Comfortable using Claude Code as part of your daily workflow
•
Able to explain complex systems simply to non‑technical stakeholders and go deep with experts
•
Track record of owning projects end‑to‑end and mentoring other engineers
Nice to have
Nice to have:
•
Distributed training (FSDP, DeepSpeed, Megatron, etc.)