Remyx - Your AI Production Assistant 💡

Plus: Remyx CEO Salma Mayorquin on the opportunities around enabling AI engineers to build faster, better and more efficiently...

January 24, 2025

CV Deep Dive

Today, we’re talking with Salma Mayorquin, Co-Founder and CEO of Remyx AI.

Remyx AI is a development workbench built to empower the next generation of AI engineers. Founded by Salma and her co-founder Terry Rodriguez, Remyx helps streamline the AI development process by reducing the heavy lifting around infrastructure, experimentation, and deployment. It enables engineers to focus on building innovative applications using the latest AI technologies, optimizing their performance, and iterating rapidly.

Today, Remyx has already started working with developer and AI teams across several industries, helping them deploy AI-powered tools like chatbots, RAG agents, and custom LLMs via its Production Assistant. By automating key parts of the machine learning lifecycle, Remyx accelerates development while enabling engineers to experiment and refine their models with ease.

In this conversation, Salma shares how Remyx came to life, the challenges of building tools for AI engineers, and her vision for how AI can unlock new possibilities for application development.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Salma 💬

Salma, welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on you and Remyx. What led you to co-found Remyx?

Hey there! My name is Salma, and I’m the CEO of Remyx AI. We’re a team building agent-guided ML-Ops. The idea is that we want to enable the experimentation process of AI application development for more folks and reduce all of the boilerplate and additional infrastructure work that engineers are bogged down with, by focusing on the experimental, iterative process of being able to work with data sources, model types, model sources, and architectural designs. What we’re building allows engineers to build the best systems and applications for the future.

How would you describe Remyx to the uninitiated developer or AI team?

Remyx is the development workbench for AI engineers. For the new wave of folks building with AI, we want to build smarter tools for them to create new applications with the latest and greatest developments in AI, improve them over time, and continue to optimize their applications. We aim to support the lifecycle of all the new, awesome apps people are going to build while reducing their workload.

Who are your users today? Who is finding the most value in what you're building at Remyx? 

The people we've been engaging with a lot this past year are often data scientists, aspiring data scientists, or adjacent engineers in fields like BI or data engineering. These folks, like many of us, are swept up in the rapid developments in AI and are trying to participate in this wave, learn new skills, or create new applications to improve their internal systems or enhance customer-facing products. They typically have some context on data processing workflows and are now tasked with figuring out how to build AI applications using LLMs, VLMs, and other emerging AI models.

These engineers are encountering the new wave of agent systems, chatbots, and other AI-driven tools. Many have experimented with tools that help with prompt engineering or deployment aspects—basically tackling parts of the complex process involved in developing AI applications. They're looking for solutions that provide them with more tools and levers to improve and continue developing these applications.

We aim to provide a workbench that not only gathers analytics automatically but also interprets them for users—a specialized co-pilot tailored to the AI experimentation process. This means capturing metrics from deployments, tracking how data sets are designed, monitoring which models have been trained, and evaluating their performance. With all that information, the platform helps synthesize actionable plans.

For example, it might identify the most promising directions to pursue out of all potential options. Users can then experiment with those paths and assess how their results align with user responses and whether they're moving closer to the desired "golden state" for their application.

Which existing use-case for Remyx has worked best? Any customer success stories you’d like to share? 

We’re still early in some of the pilots we’ve had with folks because a lot of them are still identifying which problems they want to tackle out of all the possibilities they could explore. That said, a common one we’ve seen is people with a BI analytics product or those using BI internally who want to package that query power into something more accessible, like a chatbot or a RAG agent. Many have experimented with different frameworks or systems, like deploying a custom LLM or tapping into APIs from OpenAI or Anthropic to ingest their data and expose queries for users.

We’ve been helping by identifying strategies to evaluate those questions or queries. How do you define the ideal state? How should the model perform? What are you optimizing for—precision, how cordial the chatbot is, or maybe a combination of both? We help folks understand and evaluate at different points in the process, from data and model selection to data curation, training, and deployment. We also assist with setting up some of the infrastructure.

For example, we’ll help set up a deployment to host an agent connected to a database that can write queries or link to other tools they’ve designed to answer user questions. Essentially, we help with both the experimentation process—evaluating and training the model with a dataset—and the deployment and management of the infrastructure to sustain the lifecycle of the new application they’re testing.

Walk us through Remyx's Production Assistant. What use-case should people experiment with first, and how easy is it for them to get started? 

This is where the name of our company really explains itself! Remyx is about dogfooding the capabilities we want to help others develop. Our Production Agent acts as a co-pilot that gathers information and context about your goals, the constraints your business is operating under, and how resources are allocated for your project. This could include constraints like available resources, where the application can be deployed, or other limitations shaping the development process.

The co-pilot is essentially an agent—right now it’s LLMs with tools—where you can share what you’re building, your experiences, your goals, and your constraints. Based on that, the agent gathers this context and provides a game plan template. For instance, if you need to build a chatbot that must run on one GPU for a prototype, with specific performance or accuracy requirements, the co-pilot can recommend the best choices for datasets, model selection, and how to fit your constraints while achieving the desired performance.

It also helps put everything together so you can test it yourself, gather more information, and continue refining the process by feeding that information back into your workflow.

How are you measuring the impact or results that you’re creating for your customers? 

We’ve found that a lot of folks who’ve tried creating prototypes with popular tools sometimes hit a wall—they’ve run out of ways to improve their output or take things further. Those are the kinds of people we love working with because we can help unblock them or introduce new techniques and ideas to push their experiments forward.

Another group we focus on is people spending too much time trying to figure out how to deploy their LLMs and deal with all the provisioning involved. Our tools make that process much easier. For example, if you train a model with us, we offer a click-to-deploy feature that sets up a Triton server with your own GPU on whatever infrastructure you choose—whether it’s your cloud, local setup, or something else.

We’re really focussed on automating those parts of the process so you don’t have to spend tons of time building it from scratch. That way, you can focus on designing the application and experimenting with what will actually improve it.

Given the excitement around new trends in AI such as Agents and Multimodal AI, how does this factor into your product vision for Remyx? 

We have an open-source project called VQA Synth that’s helping us dive into the multimodal space and tackle more complex model deployments. It’s based on ideas from a paper called Spatial VLM and sets up a pipeline of models to create new datasets for training BLM-type models. With coding agents getting better at handling tricky tasks, there’s a lot of potential to team up with them to build even more advanced systems.

There’s a big push right now toward creating single models that can do a bunch of things, but pairing them with specialized models to really nail specific tasks. That’s where things get exciting—combining VLMs, LLMs, and even older deep learning models to build systems designed for specific applications.

We’re focussing on helping people experiment at the system level. Sure, there are tools that can show you how one model or dataset is performing, but we want to go a step further. We want to let users test how different combinations of tools, models, and data sources work together as a system and figure application design works best.

What has been the hardest technical challenge around building Remyx into the platform it is today? 

It’s been full of challenges, and honestly, a bit of “putting our money where our mouth is.” There are so many tools out there now—agents, AI coding capabilities—and we’ve actively relied on and used those tools ourselves to expand our skills. For a bit of background, Terry and I have been ICs in the ML and data science space for the last decade. We’ve worked on a lot of parts of the stack: data engineering, cleaning and processing, model training, model deployments—you name it.

That said, while some folks might have broad experience across the board, nobody’s an expert in every single thing you’d need to architect a new product. We’ve leaned heavily on chat agents to help us with things like constructing Kubernetes deployments, which is an area we’ll need to focus on as we help people manage infrastructure in a robust and scalable way. It’s been a great way to level up and build out that kind of infrastructure.

One big challenge that still remains tricky is evaluation - figuring out how to properly evaluate models that can do all kinds of things. It’s a problem a lot of people are still trying to solve, and it’s a tough one. How do we make sure that when we’re adjusting those weights—whether it’s through prompting strategies or fine-tuning or anything else—we’re actually boosting the capabilities we care about and not messing up other parts of the model we don’t want to touch? 

That’s a big problem, and it’s something we knew early on that a lot of people were struggling with. Even just deciding where to start—like picking which model is actually the best one to experiment with—is a challenge, especially as new model weights keep getting released.

How do you see Remyx evolving over the next 6-12 months? Any specific developments that your users should be excited about? 

We’re super excited about the new coding abilities of these LLMs and baking that back into the product to expand on the agent idea. The most successful companies over the last decade—Netflix, Google, AWS, Amazon—have really robust knowledge bases of their AI strategies. They’ve built those over time through extensive A/B testing, building all kinds of datasets, training different models, and creating systems with the infrastructure to support them.

We want to get to a place where the agent can help build that kind of infrastructure and institutional knowledge that a business gathers over years of experiments, trials, and tests. Our goal is to supercharge AI engineers from the start, giving them more of an architect’s role in developing applications, rather than focusing on the lower-level, mechanical tasks of managing infrastructure.

The big vision is to keep expanding the agent with more business context and experiment history. As teams use it, the agent can learn from what teammates are experimenting with and guide the direction they might want to test. It’s about supporting that collective workflow to improve applications with less upfront investment from developers or the organization.

We want to create a copilot that’s helpful in the experimentation process, gathers intelligence, and helps people quickly improve and deploy production-grade applications.

How would you describe the culture at Remyx? Are you hiring, and what do you look for in prospective team members joining the company? 

We’re definitely looking for talent! If you love MLOps, ML development, or application development, please reach out. We’d love to hear your thoughts and  if you’re interested in joining the team.

A bit about us: we’re a team of two right now, just my co-founder Terry and m. We’ve known each other for the last decade—we met at UC Berkeley, where we were both studying math. We serendipitously found our way into the world of data science and AI when it was starting to gain traction in the data science realm. Since then, we’ve worked at a variety of places.

We also created and host our own blog called Smells Like ML, where we’ve open-sourced a bunch of cool projects. Through this, we’ve collaborated with teams at NVIDIA, ARM, and Arduino to promote different software and hardware platforms independently. Terry has worked across a variety of startups in robotics, gaming, streaming, and healthcare, while I recently finished a stint at Databricks as a solutions architect specializing in ML infrastructure. In that role, I worked with a ton of teams on their AI strategies, helping them deploy Databricks or explore other ML tools to build applications.

We’ve gained a lot of exposure to how others approach AI development beyond just our personal experience. We made the leap to Remyx because we saw an opportunity to create an awesome workbench that focuses specifically on the needs and workflows of the new AI workforce—the AI engineers building the applications of the future.

Conclusion

To stay up to date on the latest with Remyx, learn more about them here.

Read our past few Deep Dives below:

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.