Neo - Your Autonomous Machine Learning Engineer 🎛

CV Deep Dive

Today, we’re talking with Saurabh Vij and Gaurav Vij, Co-founders of Neo.

Neo is an autonomous machine learning engineer. Founded by Saurabh and Gaurav after years of working in AI infrastructure and ML development, Neo was built to solve two major bottlenecks: the high cost of building AI models and the global shortage of skilled ML engineers. 

Imagine a senior ML engineer working 24/7, handling grunt work and solving complex problems at a fraction of the cost. That’s NEO, helping companies dramatically scale their AI capabilities while cutting costs.

Unlike existing AutoML tools with their predefined recipes, NEO is a multi-agent system that can dynamically develop and adjust data engineering and model training workflows for use-cases like classification, predictive analysis, and Generative AI, turning every ML engineer into a superhuman.

Neo is already being used by high-growth startups and Fortune 500 companies to eliminate infrastructure headaches, automate tedious ML tasks, and accelerate AI development. It’s also proving to be a game-changer for software engineers transitioning into AI roles by serving as a hands-on ML co-pilot. Today, it outperforms OpenAI on the MLE Bench, winning medals in 26% of the Kaggle competitions it participated in.

In this conversation, Saurabh and Gaurav share how Neo is redefining machine learning engineering, the unique innovations behind its reasoning capabilities, and their vision for making elite AI talent accessible to every company in the world.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Saurabh & Gaurav 💬

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

Gaurav: Around eight years ago, I was running my own computer vision startup. I have eight years of experience in AI and ML development, and during that time, I ran into two major roadblocks. First, the cost of building and deploying these models at scale, and second, the scarcity of good ML talent, which came at a very high cost. As a small, bootstrapped team focused on building computer vision models, it became incredibly difficult to get these models into production for enterprise use cases. Those were the two big challenges that ultimately motivated and inspired me to work on solving this problem. That’s when Saurabh and I teamed up. 

Saurabh: I was a particle physicist with experience at CERN in Geneva, Switzerland, home to the world’s largest physics lab and the discovery of the Higgs boson. I was using neural networks and AI in physics and that experience also gave me exposure to High Performance Computing and Distributed computing. We both became obsessed with AI and computing and how we could build an AI and compute company together.

This is actually our third startup together. Our first was Q Blocks, which focused on distributed computing. The idea was simple—could we run AI models on someone else’s crypto mining rig or mining farm instead of AWS? That led to a 90% cost reduction. Then we started Monster API, where we built the world's first no-code fine-tuner, making fine-tuning models super easy. Across both of these, we served over 30,000 developers with 1 million computing hours.  This experience and deep engagement with thousands of AI/ML developers gave us deep insights into the problems everyday machine learning developers face, which ultimately led to Neo.

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

Neo is the first autonomous machine learning engineer. It automates the entire ML workflow—from data engineering to building AI models from scratch—cutting time to market from months to just a few days. Think of it as having a senior ML engineer working 24/7 for you, handling grunt work and solving complex problems at a fraction of the cost.

The best part is, it outperformed OpenAI on the MLE Bench, scoring 26% indicating performance at the level of a kaggle master

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

ML engineers, data scientists, and AI teams are our primary users, and what we have realized is they want to focus on innovation, new ideas, faster iteration, and system and architecture design. But today, they are bogged down by complex and time consuming tasks like Data cleaning and preprocessing, feature engineering, optimal model selection and hyperparameter tuning experiments and analysis, debugging pipelines, and wrestling with deployment nightmares.

Neo eliminates all that complexity, handling the tedious, difficult tasks so they can focus on what truly matters—building high-quality models and getting them to market 10 times faster. Though we started with a focus on ML engineers only, during our private beta, it turned out that many software engineers within large enterprises transitioning into AI roles are struggling with a steep learning curve, and they are finding Neo as a bridge to accelerate their journey.

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

Neo has two main components: the chat interface and the artifact section. Through the chat interface, users can provide tasks to Neo, ranging from complex ML engineering and feature engineering to data science tasks. As Neo progresses, it returns messages with logs detailing insights into its reasoning and actions it is performing, allowing users to track its progress through various minute tasks within the overall plan.

In the artifact viewer, users can see the outputs Neo generates, including model weights, pre-processed datasets, code and their pipelines. Right now, Neo excels at classical ML tasks like building recommendation engines, classification models, and predictive analysis pipelines. These types of tasks are a great starting point for users to get acquainted with Neo’s capabilities.

Which existing use-case for Neo has surprised or delighted you the most? How are you measuring the impact or results that you’re creating for your customers? 

So far, we've seen two main reactions.

First, ML engineers are thrilled that they no longer have to deal with tedious infrastructure tasks like setting up GPUs, managing CUDA, and building and orchestrating containers. In many teams, companies spend millions hiring world-class ML engineers, only for them to waste time on pipeline management, data cleaning, and data engineering instead of actual innovation. With Neo handling these tasks, engineers can now focus on system design, architecture, and building cutting-edge models.

The second major reaction is about speed. One of our early private beta customers was running 10 experiments a week. With Neo, they’re now running 60. That kind of scale leads to more variance, better model iterations, and much higher-quality outcomes. What used to take months can now be done in days.

Finally, Neo is becoming a bridge for software engineers moving into ML roles. With the rise of tools like Devin and Cursor, many software engineers are transitioning into AI, but they don’t have a decade to master machine learning. Neo acts as their personal copilot, automating the hardest parts of ML engineering and helping them level up and become proficient much faster.

There are a number of companies working on AI autonomous developers. What sets Neo apart from a user/product or technical perspective? 

What we’ve observed is that most autonomous developers today are focused on making software engineering easier, but we’re specifically focused on machine learning engineering. This focus presents a unique set of challenges.

We’re going vertically deep because ML engineering is inherently more complex than traditional software development. It’s non-deterministic and highly experimental, requiring many subjective, context-dependent decisions. Engineers need to navigate feature selection and engineering, model architecture choices, hyperparameter optimization, evaluation, and deployment optimizations with frameworks like vLLM and TensorRT.

Because of this experimental nature, ML engineering demands stronger reasoning capabilities, adaptive approach and deeper integrations with various tools. In short, our approach is about deeper specialization rather than a broad, generalized software development focus.

Could you share a little bit about how Neo actually works under the hood? 

We’re currently filing a patent, so I can’t share too much detail, but at a high level, there are three major innovations that set Neo apart and significantly enhance its reasoning and execution capabilities.

The first is multi-agent orchestration. Initially, our workflow relied on a single agent, but transitioning to a multi-agent system was a game-changer. If you’ve read Sapiens, you’ll recognize the concept—an individual human is limited, but humans who coordinate through tight feedback loops can accomplish far more. The same principle applies to AI agents. A single agent can hallucinate or provide false information, but when multiple agents—each specializing in different areas like data science, ML engineering, and data engineering—are orchestrated together in a tight feedback loop, their emergent capabilities grow exponentially.

The second innovation is fine-tuned expert models. Instead of using a single general-purpose model, we take a mixture-of-experts approach. There are around 30 tasks in a typical ML workflow, and we have fine-tuned models that collaborate with each other to implement different parts of the complete pipeline, leading to higher efficiency and less hallucination while solving the given ML problem.

The third breakthrough is our proprietary reasoning algorithm, which enhances structured decision-making, leading to improved problem-solving, learning, and adaptability in executing complex ML tasks.

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

While developing the entire system, we faced multiple challenges, many of which were among the hardest to tackle in ML engineering workflows.

One major challenge was managing the limited context length of LLMs while handling complex machine learning experiments. ML engineering tasks involve running numerous experiments with different hyperparameters, datasets, and early stopping criteria. The sheer number of variables and the non-deterministic nature of the process result in very long context requirements. Initially, fitting all this into the limited context window of LLMs was difficult, but we overcame this with our smart context orchestration and retrieval algorithm.

Another major challenge was preventing hallucinations and ensuring the system adheres to a structured plan. This is especially crucial in machine learning, where LLMs can sometimes generate incorrect or misleading results. Neo’s iterative approach reduced hallucinations significantly.

Finally, a key challenge was incorporating feedback from multiple sources and autonomously generating meaningful insights for error resolution. Debugging and course correction like a missile guidance system are critical for any autonomous agent, and we had to refine Neo extensively to ensure it could resolve errors on its own. After significant experimentation and learning, we successfully integrated this capability, making Neo a far more reliable and intelligent system.

These three—context management, reducing hallucinations, and autonomous debugging—were among the toughest hurdles we had to overcome while building Neo.

How do you view the development of open-source AI from the lens of Neo? 

Right now, we are super focused on refining and making NEO better from the feedback we are getting from our enterprise design partners, so it's a closed system for now. But at some point, we would want to open-source it because the philosophy behind the company is to actually democratize access to an AI engineer.

By 2030, the demand for ML engineering is growing so fast that there will be 1 million AI-native companies. But the problem is there are only 300,000 highly skilled ML engineers, only 5,000 Kaggle Masters, and just 600 Kaggle Grandmasters.

If you go out there to hire a Kaggle Grandmaster today, it will easily cost you anywhere from half a million to $2 million. And the reality is that most of these top engineers don’t want to work for a lot of companies. They’d rather be at places like SpaceX, Tesla, Meta, or OpenAI instead of mid-market companies or startups.

So this talent gap is massive, and it’s only growing. We want to fill that gap with an AI engineer that can do the job of an entire ML team.

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

Neo is currently fully autonomous, but based on user feedback, we see opportunities to make it more collaborative. We're designing a system that allows for partial autonomy, where users can interact more with Neo instead of relying solely on full automation.

We're also working on integrating Neo with the broader MLE application ecosystem, including third-party integrations. This means Neo won't just function as a standalone product—it will be able to connect with various data lakes, external platforms, and MLOps pipelines. This will significantly expand its capabilities, making it even more practical and usable in real-world applications.

Looking ahead, Neo already performs at the level of a Kaggle Master, but our long-term vision is to push it to the level of a Kaggle Grandmaster. Today, there are only 600 Kaggle Grandmasters in the world. Imagine if every company had access to one—think of the breakthroughs and advancements that could come from that level of AI expertise at scale.

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

We’re currently a small team, including the founders, with senior data scientists, machine learning engineers and infrastructure experts. Our team includes members with experience at Wolfram Research, and we’re also collaborating with contractors who have backgrounds at MIT and Stanford. Following our seed round, we plan to bring some of these contractors on board as full-time team members.

Anything else you'd like our readers to know about the work you’re doing at Neo?

With Neo, you can now scale your machine learning and AI team infinitely at a fraction of the cost of hiring traditional ML engineers—all while having an AI engineer that works 24/7. If you already have an ML team, Neo turns every ML engineer into a superhuman, dramatically increasing efficiency and output.

Additionally, there’s a key distinction between Neo and traditional AutoML pipelines. While AutoML is highly templatized, Neo’s advanced reasoning and auto course-correction capabilities allow it to handle complex, unforeseen ML engineering tasks. It generalizes across non-deterministic workflows, making it feel as if you’re working with a human ML engineer.

Conclusion

Stay up to date on the latest with Neo, learn more about them here.

Read our past few Deep Dives below:

• Build specialized RAG agents with Contextual AI 🌐

• Remyx - Your AI Production Assistant 💡

• Qdrant's GPU-accelerated vector indexing is here 🔋🌐

• Athena is your AI-powered remote hire 🎛

• Doowii - your AI-first education platform 📖

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.

Join Slack | All Events | Jobs