The AI and ML developers conference


The MLCon is meant to break down silos, to share lessons learned, pro tips, proven strategies from leading AI developers and data science leaders. Learn best practices and strategies in AI infrastructure, ML in production, and exciting research that you can apply to your next ML or DL project. Hear AI leaders as they share their successes, failures, and lessons learned so no one has to reinvent the wheel.

Learn from 35+ Top AI Leaders at MLCon

Get to know the speakers

Learn from top AI leaders

Michael McCourt
Maxime Bergeron

Stage 1

Yochay Ettun, CEO & Co-founder  at

Leah Kolben CTO & Co-founder at

In this talk, Arvind will discuss augmenting GPT models with retrievers. Retrievers search for relevant information that is then added to the context of GPT models. This procedure helps with factual generation improving the reliability of text generation. The same procedure also enables a way to leverage labeled training data without fine-tuning providing an autoML solution that is easy to configure, and adapt to changing label schema.

Arvind Neelakan, Open AI

Sometimes the tools we need are incredibly simple – which is great when we need them to work fast! I’ll dive in on a simple but powerful anomaly detection model that is helping us to identify when the alerts we see in engineering are having customer impact.

Margaret Campbell, Snowflake

Machine learning (ML) platforms and ML-centric systems have become a popular subcategory of software systems. They are, however, uniquely different from conventional software systems because of their close relationship with data. Data flows through these systems in various forms such as raw data, features, parameters, and predictions. This talk will try to establish – with examples from different business domains – how the quantity and quality of such data influence the efficiency and performance of an ML system. Optimizing hardware and software for ML is discussed enough, the objective of this talk is to highlight the need for optimizing data as well. Through the lifecycle of AI/ML platform development, the developers should try to understand the data for which it is being built.

Abon Chaudhuri, Robinhood

According to TechRepublic, 85% of Machine Learning projects fail. Among the reasons that contribute to this scary statistic the most prominent are lack of leadership support, strategy or engineering skills. In the talk I’ll try to examine the main pain points, explaining best practices on how to overcome these challenges, bringing on the table real world examples, personal experiences and actual insights that allowed me and the teams I worked with to successfully deploy ML at scale and drive a real business impact. More than 15% of the time.

Massimo Belloni, Bumble

The User Intelligence team in the Data Intelligence organization at CNN Digital is working on content recommendations. We are still in the experimentation phase, although that involves A/B testing with as much as 20% of traffic. Initially experimentation was incredibly slow. We had a single tenant API and struggled with the process for running experiments in Optimizely. We were focused on a small number of relatively complex models. And, we had no offline evaluation process to signal what we should A/B test. We have managed to make a ton of progress on pain points within the past year even though we still have a lot we want to improve on. In this session we will talk about what we have learned and improved and what we want to work on next

Ailish Byrne & Tim Obert

ML forms the backbone of the Lyft app and is used in diverse scenarios. While the modeling technique used by each team is different, a common platform is needed to simplify the development of these models, parallelize model training, track past training runs, visualize their performance, run the models on schedule for retraining, and deploy the trained models for serving. We built LyftLearn to achieve these goals. In this talk, we’ll discuss the high level architecture of LyftLearn, a system built on Kubernetes, which manages ML model training as well as batch predictions. We will demonstrate how we achieve: Fast iterations No restriction on modeling libraries and versions Layered-cake approach Cost visibility Ease of use

Shiraz Zaman, Vinay Kakade, Han Wang Lyft

Machine learning models are only as useful as the metrics for which they are trained and optimized towards. In this talk we’ll discuss common pitfalls when designing and tracking machine learning metrics and how to overcome them. We’ll then focus on intelligent search approaches that explore many metrics at once to better understand the modeling problem and parameter space. Finally, we’ll discuss techniques for optimizing multiple metrics at once to analyze tradeoffs. This talk will provide useful lessons for developers just getting started in ML, engineers fine-tuning pre-trained models for production, or seasoned researchers developing and training algorithms from scratch.

Scott Clark, SigOpt

Data science and analytics has moved from being an investment in the future to a core component of corporate strategy. In the rush to stand up this new practice, many organizations have had struggles in realizing value. This presentation would provide insights into how to set up an effective practice by focusing on People, Process, and Strategy, and use the specific case of WestJet’s recent reboot of their analytics function. This is based on an upcoming book by the same name.

Jeremy Adamson, WestJet

Data Science is a vast discipline with research professionals and brilliant scientists working on cutting edge AI and ML technologies. But translating the impact of a model on revenue and margins to generate business value is essential for success. To achieve this, a generalist approach is required, where professionals can think beyond the models and algorithms and understand that data is an enabler in a vast scheme of things. More often than not simple, understandable models solve real-world problems as they are robust, scalable and readily trusted by conventional teams.

Mousami Mishra, Vodafone

Kubernetes has become de facto the operating system for running workloads over the cloud. But, the question is whether it is also the best tool for AI workloads. In we have been running AI workloads in Kubernetes for many years, and we have learned many lessons. In this session, Itay Ariel, backend team lead will examine the existing solutions for running AI workloads over Kubernetes, specifically using native Kubernetes solutions, Kubeflow, volcano and in house solutions.

Itay Ariel,

Computer vision applications in the medical domain have recently become quite popular. Detecting, localizing, and diagnosing diseases and recognizing structures on MRI, CT, PET, XRay, ultrasound and photographic images is efficiently done by AI. What is rarely discussed, is how these AI systems are made. They require copious amounts of training data consisting of images and human-supplied labels. The labels are marked areas (either rectangles or free-form boundaries) that are attached to a word, e.g. “tumor.” Naturally, only highly qualified professionals are able to provide these labels making the process effortful and expensive. A technique called “active learning” from AI can help with this by reducing the manual effort by 90%. This allows the creation of state-of-the-art AI models for medicine using a significantly smaller budget of time and resources. This talk will present the method with several examples and will argue that this approach is a disruptive shift in medical AI.

Patrick Bangert, Samsung SDS

Human behavior is dynamic. From mood swings throughout the day to adopting different habits for each day of the week, time is an essential part of understanding changes in behavior. In this talk, we discuss the effect of time in machine learning systems that require understanding user behavior. We present ideas for keeping systems updated over time, methods to leverage time as a way to improve training data, and share lessons learned from studies and experiments conducted by Snap Research on the Snapchat platform.

Leonardo Neves, Snap Inc

Stage 2

Leading organizations are successfully deploying machine learning into production to innovate, grow revenue, and reduce cost. However, the path to ML is fraught with both existing and new challenges, from access to scalable data to creating operational procedures that support repeated, real-world deployment. In this discussion, Jeff Sternberg and James Tromans from Google Cloud’s CTO Office will discuss what is working, and what isn’t, and explore how CTOs and technology leaders can improve our collective abilities to leverage ML.

Jeff Sternberg, James Tromans, Google

Running Deep Learning algorithms on low-memory low-compute devices is a challenging but often required task. We developed a Deep RL algorithm for the task of optimizing datacenter Congestion Control. In this talk, we will discuss the process of deploying a Deep Learning algorithm inside a Network Interface Controller (NIC), satisfying inherent memory and computational constraints. More specifically, we will discuss the algorithm’s quantization method, writing deep networks on native C, and the methods we used to reduce the model memory consumption while keeping operation precision and efficiency.

Benjamin Fuhrer & Doron Haritan, Nvidia

AI & ML has been remarkably successful in many industries and domains, so it is only natural that the pharmaceutical industry is adopting data-centric approaches at speed. However, the nature of drug development means that it far from ‘just’ another data problem. Here I outline it’s unique aspects that raise challenges to AI and require adept use of data, ML technology and subject expertise.

Paul Agapow, AstraZeneca

Marks-and-Spencer is a UK based retailer that is undertaking a transition from an ‘old-school’ bricks-and-mortar enterprise to a digital-first company. As part of that transition, an enterprise data science team was established 1 year ago – with the remit of deploying ML to improve decision making within the business. In this talk I am going to cover our journey as we targeted and executed our first ML use cases, the challenges and learnings from building business stakeholder trust as well as the pain points we experienced moving our initial use case to production. I’ll also cover observations on the most effective ways of working and team structures that we have found for establishing a successful data science function.

Rushen Patel, Marcs & Spencer

If it looks like a duck, swims like a duck, and quacks like a duck – It’s a K8s without containers! Find out why to use K8s without nodes and containers, and what problems such a unique K8s can solve in your machine learning workflow. What will K8s look like without containers? Or Without nodes, without CNIs or storage provisioners? In this session I’ll share with you how we, at cnvrg, are building K8s style, declarative APIs for running AI/HPC workloads on bare metal by prototyping K8s API server with kcp-dev project.

Dmitry Kartsev,

The establishment of “Enterprise Knowledge Graphs” have been steadily on the rise. Despite many precedents, the pharmaceutical industry in general has been lethargic towards the implementation of such knowledge bases, even though the promise of it has been quite tantalizing. The pharmaceutical and healthcare industry generates massive amounts of data, yet they are often siloed, thus preventing the utilization of their inherent connectedness towards providing more holistic information for caregivers and ultimately providing better quality of life for patients. Here, we discuss a strategy for building a insight generation engine and a semantic enterprise scale knowledge graph, and the utilization of these to radically transform the messaging of the pharmaceutical brands. Exploiting the relationship between the various data sources, the content that defines a drug “brand” can be dynamically generated to better inform patients and caregivers about which part of their unmet needs are better served by the brand in question. By utilising the power of power of personalised messaging, the overall aim is to ensure the application of right treatment paradigms at the right time to improve disease prognosis and thus providing improved quality of life.

Srayanta Mukherjee, Novartis

In recent years, increasingly large Transformer-based models such as BERT have demonstrated remarkable state-of-the-art (SoTA) performance in many NLP tasks. However, these models are highly inefficient and require massive computational resources and large amounts of data for training and deploying. As a result, the scalability and deployment of NLP-based systems across the industry is severely hindered. In this talk Ill present few methods to efficiently deployed NLP in production, among them Quantization, Sparsity and Distillation.

Moshe Wasserblat, Intel

Need to serve a large, diverse community of data scientists, developers, and/or researchers who all need access to their own unique AI/ML tools, frameworks, and datasets? Come hear about how St Jude Children’s Hospital and their partner, Mark III Systems, worked together with to run a successful pilot in the datacenter around a multitude of AI/ML use cases and data science users. We’ll share insights about how to plan and architect for a successful pilot, tips for success, and lessons learned.

Chris Bogan & Dr Franz Parkins, St Judes & Mark III

Building business and consumer facing NLP platforms and systems at scale for high load, many models, high business results and consumer satisfaction

Andrei Lopatenko, Zillow

Adopting data science could potentially advance and accelerate business growth, yet it has proven to be not so trivial across all industries. Time after time, it has shown that merely collecting data and hiring a team of data scientists are not sufficient enough. What are the important ingredients in creating a sustainable environment so that you can leverage the data scientists skills to their full potential and keep them engaged?

Maria Lee, Roche

Many researchers have attempted to measure the respective effort that data scientists expend on preparing data for modeling vs the time spent training and evaluating candidate models. The results have been surprisingly consistent with most estimates for data prep being reported as approximately 80% of the total analysis time. However, the skills and tooling requirements for data preparation, especially for distributed systems, are not getting as much attention as the less time-consuming modeling phase. This talk looks at options for distributed data preparation that allow data scientists to experiment with data pipelines and still have time to focus on modeling.

Philip Hummel, Dell

Product search is a key functionality for most e-commerce platforms. Zulily uniquely marries personalization and discovery shopping while also serving the customer need to search. This means that getting search right while maintaining the discovery aspect requires a unique approach that uses large amounts of detailed product information along with behavioral data from customers to predict what customers want. In this talk I will present some general approaches Zulily uses to improve search relevance. Along the way I will pay special attention to areas where ML can help with this process and highlight where a robust ML platform is essential.

Tyan Hynes, Zulily

Stage 3

Attention is a valuable resource for rapidly scaling companies; the time it takes to manually monitor dashboards for new business trends can be crippling to new initiatives. Gazer is Wish’s general solution for teams to build intelligent dashboards and customize email alerts. With millions of combinations of data segments and metrics, anomalies are almost guaranteed to be found; so the primary problem to be solved is how to rank anomalies, with a goal of recommending the most useful and concise pieces of information to stakeholders without missing anything important. I’ll discuss how we approached this problem at Wish, what we’ve learned along the way, and how one might improve upon our current design.

Chandler Phelps, Wish

ML Platforms are hard to get right. It isn’t uncommon for custom MLPs to commit various design sins, like false prescription, that make them hard to use or limited. In some cases, the ML life cycle has done more harm than good, focusing engineering teams on common activities instead of common computing abstractions. Leveraging existing systems principals, we propose a possible ML Systems layered approach. As a tangible example, we focus on data versioning, examples of which exist across commercial and private MLPs. We describe our experiences developing and using Disdat, an open-sourced data versioning system, to make the case for interoperable ML systems that can accommodate complexity and innovation.

Ken Yocum, Intuit

When every app or website has a search box, it is crucial to level up your search engine to rise past the competition and increase your bottom line. Introducing machine learning to your search engine is complicated but worth it – so let’s break it down. This talk details the transition from a simple weight-based search engine to a machine learning powered, data driven setup; and how to achieve the „Always be testing“ state with rapid iteration cycles. It covers an end-to-end search engine architecture, from data logging from Microservices, processing with Apache Spark, training with LambdaMART, and deploying your models with ONNX on top of ElasticSearch or Solr.

Samay Kapadia, Delivery Hero

Visual Search is becoming more and more important across the board but especially in ecommerce. Many companies are beginning their own inhouse products, such as Amazon’s StyleSnap and seeing huge upticks in usage. At ShopRunner, a company that allows our members to shop our 150+ retail partners and receive benefits when they do, we have created a scalable visual search platform that customers, and the internal marketing team, can use. This gives anyone the ability to quickly find visually similar products in our catalog of 3milllion+ active options. In this talk I will walk through our path to get there and how our current deployment system is set up with ElasticSearch.

Morgan Cundiff, ShopRunner

Retail, Campuses, Manufacturing, Transportation, Airports, Public Areas, Critical Facility & Physical Security are the building blocks for Smart Cities. 3D Lidar is an ideal solution for gathering business, operational and safety insights, and alerts in a variety of environments, without capturing private information. They help all areas and organizations become safer and more effective so they can thrive. With 3D Lidar, organizations and cities can gain detailed insights and utilize data-driven solutions to:
• Keep traffic flowing and transit safe, with integrated video, data, and alerts.
• Protect students on campus while achieving operational excellence.
• Provide downtown area insights on foot traffic, parking, traffic mix, and more.
• Drive more retail success with business intelligence and insights for a thriving
Hitachi leverages IoT, video, lidar and data management solutions that help our customers reach the out-comes they seek.

Aniket Patange, Hitachi

Habana® Gaudi® is the industry’s first AI training processor to natively integrate ten 100-Gigabit Ethernet RoCE ports on-chip, enabling flexibility of scaling up and scaling out using industry standard interfaces. Gaudi’s compute efficiency and integration brings new levels of price-performance to both Cloud and data-center customers.  Riskfuel, a fintech startup, is pioneering the use of deep neural networks to learn the complex pricing functions used to value over-the-counter derivatives.  At re:Invent 2020, AWS announced that “the new AWS Gaudi-based EC2 instances will leverage up to 8 Gaudi accelerators and deliver up to 40% better price/performance than current GPU-based EC2 instances for training DL models.”  The significant cost-performance advantage makes Habana Gaudi a compelling addition to RiskFuel’s AI compute infrastructure.  Habana’s SynapseAI software stack is integrated with TensorFlow and PyTorch frameworks, and it takes a few lines of code to get started with Gaudi.  In this talk, we will share our learnings in migrating deep learning models to Habana Gaudi to reap the cost-performance benefits.



Sree Ganesan, AI SW Product Management at Habana Labs

Maxime Bergeron, Director of Research & Development at Riskfuel

You’d think APIs are simple. You send a request and get back a response, what’s so complicated about that? But sometimes the response takes too long to compute. Google recommends you aim for a response time lower than 200 milliseconds, everything over half a second is an issue. For example If you developed a Deep Learning algorithm and you want to share it with the world, you need to develop an API exposing it. But you obviously can’t compute the algorithm for 10 minutes before returning a response from the API. No user will wait that long, and you most certainly should not use an expensive GPU that has a great compute power in order to serve the API requests. In this talk we’ll explore tools for solving this problem and building a scalable system. We will get to know Celery, which is an open-source, asynchronous, distributed task queue. It will save you blood, sweat and tears when trying to set up a distributed workers system to perform tasks for your API.

Nir Orman, Wix

In this presentation, I explain how we develop a sentiment analysis model using Bert-transformer. This sentiment model is developed to classify employee’s comments from different surveys within ING. We use a novel method to compare model’s outputs continuously in order to monitor the model’s performance. Simultaneously, by using active learning, the algorithm proactively flags comments which is not confident enough for labeling. Therefore, this method ensure that the annotators only annotates the most important and difficult comments, thus making the whole process more efficient and boost model performance.

Mojtaba Farmanbar, ING

Salesforce Datorama platform is highly customizable, allowing hundreds of different data connectors mapped into a unique marketing data warehouse. This presents great value for our customers, since they can leverage the platform to their own specific needs, however this also creates complexity when the variety of data increases. This is where ML comes to rescue – we will show how we are helping our customers scale and manage their own data, using on-line, multi-tenant training.

Amir Kafri, Salesforce

Content is at the heart of what Adobe does. While most of this content gets used in marketing, its richness in marketing AI applications is relatively less explored. To bridge this gap, Adobe’s Digital Experience business embarked on a journey in building a suite of AI services, called Content and Commerce AI. At a high level, these services are aimed at extracting intelligence from content, either text or images. While one could extract key phrases, entities, sentiment among other things from text, label them per customer defined taxonomy, and more. Likewise, one could extract color profile, objects, faces, text from images, classify them per customer taxonomy, and more. Such metadata could then be leveraged to improve document search, recommendations, building visitor interest model among other applications. In this talk I will talk about various services that are built as part of Content and Commerce AI, including text services such as key phrase, entity and concept extraction, relevance ranking, text classification, image services such as color extraction, image classification, OCR, face detection and recognition, conditional generative models for image generation. Finally as a overarching theme, I will talk about the problem of building AI models as a service for enterprise customers. This will be a beginner level talk where I will introduce the audience to the use-case and problems, with a brief mention of our solution approach. The audience will get to understand the problem space at a high level and possibilities that such solutions can open up. They could use this learning to apply the same in their respective organizations or leverage Adobe solutions for the same.

Deepak Pai, Adobe

Training data is critical to create robust ML/DL algorithms. Some researchers and visioners think that data is even more important than the algorithms themselves. In this talk you will learn some best known methods to deal with training data, understand problems with data collection, realize difficulties with data annotation, and find out how some Intel teams manage data in real projects. Also you will get a short overview of tools which are developed by Intel like CVAT ( and Datumaro ( for annotating and managing training data.

Nikita Manovich, Intel

Learn how to build an end to end NLP pipeline with BERT in PyTorch. Vasilis will show you how to move from research to production and implement an NLP pipeline quickly and efficiently using PyTorch and to deploy a BERT Question and Answer Bot. In this session Vasilis share best practices for building your NLP pipeline, and how to create a seamless, reproducible workflow. You’ll follow an end to end example that will help you solve your next NLP problem, and strategies to maintain your model in production.

Vasilis Vagias,



Join the innovative enterprises using to build, train & deploy their models faster.

schedule a demo today

Announcing CORE, a free ML Platform for the community to help data scientists focus more on data science and less on technical complexity

Download cnrvg CORE for Free

By submitting this form, I agree to’s
privacy policy and terms of service.