If I was hiring an AI Engineer position here is what I would look for in an ideal candidate:
- Strong knowledge of Python & OOP
- Able to work with Pandas to manipulate data from various sources and produce train, validation, test set splits using off-the-shelf libraries
- Can identify basic data leakage issues
- Proficient in Linux and terminal usage
- Able to write unit and integration tests and apply mocking
- Experience with version control, branching, pull requests, code review in a team setting
- Ability to configure Github Actions / Azure Dev Ops pipelines to perform automated testing and deployment
- Ability to leverage Docker to containerize applications and dependencies
- Able to leverage off-the-shelf machine learning frameworks (scikit-learn, HuggingFace Transformers) to train models for common AI/ML tasks
- Working knowledge of which metrics apply to which tasks
- Ability to work with the OpenAI API in Python and write basic code structure for handling message formatting, conversation loops, function calling
- Able to apply the concepts of templating (Jinja2), unit/integration testing, and mocking to Large Language Model Applications
- Able to build demo UIs for AI applications using Gradio
- Learn how to leverage Large Language Models for traditional ML tasks such as classification, regression, & text generation and how to adapt LLM solutions such that their performance can be measured with common ML metrics
- Articulate the pitfalls of utilizing text metrics such as edit distance, ROUGE, & BLEU
- Be able to apply in-context learning to solve problems with large language models
- Able to utilize frameworks such as Axolotl to adapt pre-trained large language models using continued pretraining, LoRA, & QLoRa
- Able to estimate GPU memory requirements for training LLMs based on token count and parameters
- Able to code a basic training loop using PyTorch utilizing autograd to solve
- Able to use HuggingFace transformers to leverage the ecosystem of existing pretrained LLMs and architectures and adapt them to new tasks either via continue pretraining, LoRA, & QLoRa but also via transfer learning
I'm interested