Who Are You? The E2E RAG Stack Man!

The Real E2E RAG Stack // Sam Bean // MLOps Podcast #217

This was my first time meeting a competitive pinball player! So you could say Sam's not just an MLOps Wizard; he's a [COPYRIGHT].

While trying to follow the most up-to-date system design, there can be a tendency to lose track of your goal and end up bouncing from bumper to bumper, but Sam is very pragmatic in his approach. He discusses the importance of starting with simple metrics and basic techniques and gradually transitioning to more complex systems.

He also advocates using a core of human-labeled and synthetic data for model improvement, the relevance and potential of multimodal RAG, the challenges of creating and maintaining search systems, and the necessity of human feedback.

Plenty for my generation to think about so we won’t get fooled again.

Video || Spotify || Apple

Join us for a discussion with Anjali Balagopal, Rad AI Senior Machine Learning Engineer, for a session geared for ML professionals on how Rad AI reduced model failures and boosted AI/ML testing efficiency.

You’ll gain insights on:

Understanding Common Pitfalls: Delve into the frequent causes of model failures in ML projects and learn how to anticipate and mitigate these issues early in development.
Efficient Testing Methodologies: Discover cutting-edge testing methodologies that can significantly reduce your time-to-market while ensuring your model’s reliability and performance.
Best Practices in Model Development: Gain insights into the best practices for ML model development, including data preprocessing, feature selection, and model selection, to enhance your model’s accuracy and efficiency.

A Decade of AI Safety and Trust // Petar Tsankov // MLOps Podcast #218

When Petar told me they’d opened a second office in Bulgaria, I thought, wow, they’re doing really well; he’s bought an island in the Black Sea!

It turns out it’s a different Peter, but it was great to hear how they’re branching out into Eastern Europe and the setup they've got with the research labs.

We discussed how to bulletproof your system, and he emphasized identifying the blind spots. Knowing where the model struggles allows for data improvements to prevent future errors. We also chatted about the wider issues of AI deployment, like responsible use, potential risks, certifications, and regulations.

It was a great chat, and I’m looking forward to having him on again. We’ll record from his private island!

Video || Spotify || Apple

💡Job of the week

Machine Learning Engineer // Truveta (US, remote)

Truveta wants to enable researchers to find cures faster, empower every clinician to be an expert, and help families make the most informed decisions about their care.
Responsibilities:

Collaborate on designing and refining generative models.
Leverage expertise in machine learning and natural language processing.
Develop, train, and optimize large language and GPT-like models.
Stay updated on advancements in language and generative modeling.

Requirements:

Strong background in NLP and familiarity with advanced LLM architectures.
Proven leadership abilities in guiding a team of ML engineers and strong collaborative communication skills.
Advanced degree (Ph.D. preferred) in a relevant field, with specialized research in machine learning, NLP, and generative AI.
Proficiency in Python and deep learning frameworks (e.g., PyTorch, TensorFlow), with a knack for writing clean, efficient code.

State-of-the-art Open Source LLMs, Fine Tuning & Other Things // Luke Marsden // IRL #68 Bristol

*Assume Dick Van Dyke/Don Cheadle accent*
Owright me old China plate, over to Blighty, for a lora info in this IRL, including an intro to the idea of "quantized Lora".

Right, stop that, it's silly. And not even a Bristol accent. This talk by the LLM (that's, Legendary Luke Marsden) looks at using "quantized Lora" for making fine-tuning more memory efficient, using Lora files to enhance language models and expanding use cases for fine-tuning models. The talk also looks at the downsides of quantization, the implications of AI in various industries, some thoughts on the promise of multimodal models, and an exciting new project, Helix.

Watch it here!

MLflow On AWS With Pulumi: A Step By Step Guide

Do you feel like you're swimming against the flow, trying to keep track of many experiments and their results?

This blog is a tutorial on deploying an MLflow tracking server on AWS with Pulumi. It covers installations, setting up AWS services, security configurations, and launching the MLflow server—it's all there!

And all aimed at enhancing reproducibility, scalability, and teamwork in data science workflows. You'll definitely see results if you experiment with reading this blog!

With thanks to Bojan Jakimovski for their contribution.