improbable - AI/ML Engineer
Requirements
• 5+ years building production Python systems (backend services, APIs, data processing) • Strong software engineering fundamentals: design patterns, testing, debugging, profiling • Experience integrating LLMs into applications (OpenAI/Anthropic APIs, prompt engineering, streaming, PydanticAI) • Understanding of ML training workflows (even if you're not an expert. You need to know enough to build the infrastructure) • Docker, CI/CD, production deployment experience • Can read and understand PyTorch code (you don't need to write novel architectures) • Fine-tuning experience (LoRA, full fine-tuning, QLoRA) • Distributed training basics (DeepSpeed, FSDP) • Graph databases (Memgraph, Neo4j) • Supply chain or logistics domain knowledge • Experience with agent frameworks (LangChain, PydanticAI, etc.) • What you'll work with • Backend Stack: Python, FastAPI, PydanticAI, FastMCP, Memgraph, Postgres • ML Stack: PyTorch, Unsloth/Axolotl for training, vLLM for inference, Weights & Biases • Models: Qwen 2.5, Llama 3.1, GPT-4, Claude (for now) • Infrastructure: AWS (flexible), Docker, Kubernetes, GPUs when needed • Team: Principal Engineer (your partner on architecture), Mid Data/ML Engineer (your data pipeline partner), Junior AI Engineer (your mentee) • Example projects you'll own • Build a FastAPI service that handles streaming LLM responses with correct error handling and retry logic • Create a training pipeline that processes production logs, validates data quality, and triggers fine-tuning runs • Deploy a fine-tuned 7B model with vLLM that beats GPT-4 latency while maintaining quality on our domain • Design the data ingestion architecture for Project Genome, how we process papers, documentation, and operational data at scale • Implement evaluation frameworks that catch model regressions before they reach production
Benefits
• You're a strong production Python engineer: You write clean, maintainable, tested code. You understand async/await, know when to use generators vs lists, can profile performance bottlenecks. You've built FastAPI services (or similar) that handle production traffic. Your code passes review without drama. • You've built with LLMs in production: You've integrated GPT-4/Claude into real applications, handled streaming responses, dealt with rate limits and retries, cached intelligently. You know the practical challenges: prompt engineering, context management, error handling, cost control. • You've trained or fine-tuned models: Whether it's fine-tuning LLMs, training classifiers, or running experiments, you understand the workflow. You've dealt with training data quality, evaluation metrics, and overfitting. You can debug why a model isn't learning what you expected. • You think like a systems engineer: You design for failure, add instrumentation, consider edge cases. You know that "the model works on my laptop" isn't shipping. You care about monitoring, logging, alerting, and graceful degradation. • You can navigate the ML landscape pragmatically: You know enough about transformers, attention mechanisms, and training dynamics to make informed decisions. But you're not precious about it. If a simple heuristic beats a complex model, you ship the heuristic. • You balance velocity with quality: You ship incrementally and iterate based on production data. But you don't accumulate tech debt, you refactor proactively, write tests that matter, and leave the codebase better than you found it. • You communicate trade-offs clearly: You can explain to the team why we're choosing LoRA over full fine-tuning, why we're deploying on Fireworks instead of self-hosting, or why a 7B model might beat a 70B model. You help everyone make informed decisions.
Apply in one click
Upload My Resume
Drop here or click to browse · Tap to choose · PDF, DOCX, DOC, RTF, TXT