Revolutionizing AI with Our Vision for User-Friendly AutoML: No Code, No Data, No Problem

Try the Remyx Engine and create your custom model today!

We can expect AI adoption to accelerate over the coming decade. Still, not everyone will become an AI engineer. So the demand for putting ML into production far outstrips the talent supply. Significant gains in productivity through tools like github’s Copilot only make the bottlenecks of the ML workflow more acutely felt by the developer. And so, with the rise of no-code platforms, more developers are turning to autoML solutions to support their next ML-enabled feature.

The increasing ease of application development will bring 100X more people into the game. We can anticipate weekend hackathons that result in fully productized AI-powered solutions. With the estimated 4X increased productivity experienced by many developers using code generation today, development cadences like 12 products in 12 months will become leisure. As the cost of custom computer vision drops, mobile app developers can create AI applications on a weekly or even daily cadence.

The demand for tailor-made machine learning models is snowballing as more businesses implement ML-powered product features. Product managers can quickly test their ideas and iterate with no-code tools before committing high-cost specialist hours.

Custom ML models can be optimized for both speed and accuracy, easily outperforming off-the-shelf pre-trained models, which are prone to false positives by making irrelevant predictions. If you don’t expect to encounter trains and planes in deployment then why would you just run a detector trained on benchmark datasets like VOC?

An often overlooked advantage in fine-tuning is that custom models make it harder for attackers to exploit them than openly available models. Put another way, publicly available models are susceptible to adversarial attacks, whereby a malicious user takes advantage of your service by crafting a data payload to exploit known model vulnerabilities.

Despite the apparent advantages of specializing your model, creating and maintaining custom models has been difficult, requiring considerable expertise and resources. The process typically involves:

• Data Collection and Preparation: Acquiring relevant data for custom models can be a significant hurdle. Data must be structured before training, and the associated time and labor costs can be prohibitive.
• Model Selection and Design: Identifying the optimal model architecture and hyperparameters for your use case demands a deep understanding of various ML algorithms, their strengths and weaknesses, and the trade-offs between speed, accuracy, and complexity.
• Training and Validation: Custom models must be trained, which can be computationally expensive and time-consuming, especially when dealing with large datasets or complex training strategies.
• Deployment and Optimization: After developing your custom model, it must be optimized for inference on your target hardware, overcoming technical challenges such as compatibility, optimization, and versioning.

With all these moving parts, it’s no wonder AI projects are so prone to failure.

To address these challenges, we bake best practices into the defaults and harness the latest advances in machine learning to make custom ML model development easier and accessible to a 100X larger audience of builders.

No Data, No Problem

Traditional AutoML tools require users to curate datasets by spending considerable time annotating them on their platform. But this, too, can be automated! Ahead of the research, we have developed a multi-modal image generation and retrieval pipeline to augment your data with authentic and synthetic samples. Our approach addresses the bottlenecks in model development by automating dataset design.

Since we design the layout of your data, ground truth is free, so give your mouse a break! We’ve got pixel control to adapt datasets to your specific deployment requirements on-the-fly. We’re on a mission to make the best, most straightforward, fastest custom models on demand to support your ad hoc vision tasks optimized for our ad hoc world.

Industry reports indicate that 80% of data science efforts concentrate on data preparation, which assumes you have data access. The data collection, curation, and preparation process is a significant proportion of the average 1-6 months it takes to deploy a machine learning model. The innovator’s cold-start problem means a suitable training dataset may not be accessible.

That’s why we’ve bundled a solution for computer vision with no code or data required to create custom models. Filling out our simple natural text form takes four clicks to produce a model in minutes.

By pairing image generators with traditional autoML techniques, we’ve lowered the barrier to creating custom vision applications. We’re focused on robust computer vision, which works without demanding the end user meet us in a data scientist or engineer’s workflow.

With years of experience innovating custom AI applications, we’ve had to hone our skills in few-shot learning. We’re bullish on this technique and plan to push the limits of limiting user requirements to get started with custom ML.

Emphasizing “Auto” in AutoML

We aim to enable everyone to create and deploy their own custom models. An end-to-end, opinionated AutoML platform will guide users through the design process, meeting their specific problem requirements without needing expert knowledge. But we also find experts abstracting away simple tasks and iterating faster on new vision pipelines with automated workflows. Describe your task using natural language, and we’ll handle the rest.

We also anticipate the trend to automate more workflows with generative agents. Generative agents will perform more complex work algorithmically by iteratively querying LLMs using well-crafted prompts. But because of the tendency for LLMs to hallucinate in generation, specific knowledge-based tasks require external memory or the use of APIs. We expect efficient agents of the future to rely on custom vision models optimized for ad hoc bulk image analysis. Likewise, automated model creation can support semantic queries on image databases by precomputing indices to add structure to your data lake.

A consequence of our emphasis on ease of use is that our platform is so simple that an agent can use it through our API!

Simple and Open AutoML for Humans

The models generated by the Remyx Engine depend on widely-used deep learning frameworks and feature model architectures designed for portability. Our commitment to open-source principles ensures you are not locked into proprietary formats, promoting collaboration and innovation in the machine learning ecosystem. We’ve open sourced Remyx models shared in our model zoo and paired them with working examples to get you started.

So what are you waiting for? Sign up for Remyx Engine compute credits and create your custom vision app in minutes!