What AI and data science projects have you worked on that are innovative?

At Voxel51, our team built VoxelGPT, a chatbot for chatting with and better understanding your computer vision data. While the patterns in tabular data like SQL are relatively straightforward for an LLM to understand and mimic, the unstructured data encountered in computer vision applications is much more challenging! It took a ton of prompt engineering , retrieval augmented generation, and validation to get VoxelGPT into production!

In my experience, while developing chatbots, it's not just about coding responses or understanding intent. A significant part of creating a functional chatbot is understanding human psychology and behaviour. By integrating sentiment analysis into chatbots, I have seen them shift from mere transactional agents to interactive ones that can adapt their responses based on the emotional tone of the user. A student of mine once developed a mental health chatbot for a research project. It was programmed to provide coping mechanisms to users, especially in instances of stress or anxiety. It's innovations like these that take AI beyond its technical realm, enabling it to genuinely impact human lives.

Also, try to create the chatbot relating to a real world import use case. For example: your chatbot can act like a food nutrition expert, or daily exercise or yoga planner. This not only demonstrates your ability to develop something, but also helps you to think to solve real problems, and demonstrate your thinking ability.

The mentioned AI/ML capabilities are not mutually exclusive. E.g., we did some interesting work leveraging Deep Reinforcement Learning (RL) to (self-)improve Chatbots. The RL model leverageed data from user conversations to improve chatbot performance. At the core of our approach was a score model, which was trained to score chatbot utterance-response tuples based on user feedback - used as rewards for the RL agent. Policy learning took place offline, implemented using a DQN agent with epsilon-greedy exploration, tailored to effectively include fallback answers for out-of-scope questions. We showed an increase in performance for an Enterprise Chatbot from an initial 50% success rate to 75% in 20–30 training epochs.

A bot that pronounces names. Built a lot of chatbots and the ones I’m most proud of are the ones built for social impact. But one remains close to heart and this one’s probably the first one - a bot that could pronounce the name of a person. The bot used custom neural voice and accounted for region/dialect- providing even phonetic pronunciation. Used AI and Gen AI services.

I am yet to find an Image Generation App that can easily generate human image naturally with simple prompt. Often the work need to be specific, long prompt and complicated to get simple visual esp human image done.

The YOLO model is primarily designed for computer vision tasks related to object detection. It involves identifying and localizing objects within an image swiftly. YOLOv8, a variant of YOLO, is specifically engineered to be fast, accurate, and user-friendly. This makes it an excellent choice for a diverse array of object detection and tracking applications. Its versatility extends to tasks such as segmentation, image classification, and pose estimation, making it a robust solution for various scenarios.

Before diving into applications of image detection and generation, we must prioritize diversifying our datasets, especially in areas like facial recognition. Biases in data can lead to biased AI outputs, as seen in generative AI producing images of people in certain professions with racial biases. While innovation in this field is expected to grow, we should remain cautious, as unchecked applications may amplify existing inequalities in our world.

It is indeed an exciting frontier in AI and data science. It represents the intersection of cutting-edge technology and human creativity. In this domain, advanced machine learning techniques, including neural networks and generative adversarial networks (GANs), play a pivotal role in enabling computers to understand and generate visual content. These models, powered by vast datasets, have the potential to revolutionize industries ranging from healthcare to entertainment. Moreover, the ability to create projects that can discern intricate details in images or generate realistic visuals from abstract inputs showcases the depth of understanding and proficiency in machine learning.

Our project aims to streamline the process of interacting with SQL databases by providing a dual-functionality tool. First, it automates the generation of SQL queries based on user inputs, significantly reducing the time and effort required to construct complex database queries. Second, it offers intuitive data visualization features, allowing users to explore and analyze their SQL query results through interactive visualizations. By combining query generation and data visualization into a single platform, our project empowers users to efficiently extract insights from their databases with ease.

I've constantly leaned on data visualization for its unmatched power to articulate complex narratives. I once worked on a project visualizing the nuances in customer sentiments towards a product launch, utilizing Python libraries like Seaborn and Plotly. Rather than a table of numbers, stakeholders were presented with an interactive heatmap, revealing patterns unnoticeable before. This immediate visual feedback bridged the gap between raw data and actionable insights, underscoring the need for apt storytelling. As we forge ahead in this data-driven world, mastering this blend of analytics and visualization remains paramount.

one thing I found useful when exploring data visualization and storytelling techniques was the concept of "show, don't tell." Rather than simply telling the audience insights from the data, effective visualizations aim to highlight patterns, relationships and trends in a way that reveals the story inherently within the data itself. For example, a visualization that overlays time series data on relevant historical events allows viewers to directly observe correlations and causal relationships instead of being told about them. This forces the audience to draw their own conclusions from the data, making for a more engaging and persuasive presentation.

Data visualization in the context of personal development can go beyond simple dashboards. At Next AI Labs, we're developing features for IXCoach that will leverage Python-based analytics to elevate the coaching experience. While we currently offer a multitude of coaching styles and personalized roadmaps, our future iterations aim to utilize real-time data visualizations. These aren't just about tracking progress; they're part of a larger framework designed to make our AI algorithms increasingly effective. As we integrate these advanced analytics, users will not only understand their journey better but also experience more targeted and adaptive coaching.

Na Educa??o corporativa a IA se tornou uma ferramenta fundamental de apoio ao desenvolvimento de colaboradores, auxiliando na: - Produ??o de Texto e Síntese - Constru??o de Jornadas de Aprendizagem personalizadas - Dimensionamento de nível de conhecimento x oferta de conteúdo - produ??o de conteudo Para aqueles que tem conhecimento, foi um facilitador na constru??o de Storytelling

Automatic web content tagging for SEO and recommendation purposes has always been a key problem for content management systems. The requirement is to have a dynamic taxonomy that reflects the content at one particular point in time, but that is also relatively stable across time to allow for Intent calculations to remain sensible over time. Generative clustering based on parametric distributions such as LDA provided the stability, but lacked succinct and meaningful topic descriptions. Adding an LLM chatbot to the equation to provide topic descriptions solves this problem. This is a relatively small additional engineering step that adds significant value.

In my experience, working on recommendation systems has been both fascinating and rewarding. Let me share 3 important insights: The Power of Data: At its core, recommendation systems rely heavily on data. The more accurate and comprehensive your data, the better your recommendations will be. A/B Testing: Rigorous A/B testing is your compass. It helps you quantify the impact of your recommendations on key metrics like engagement, conversion rates, and revenue. It’s how you measure success. Lastly, remember that the true magic of recommendation systems lies in making users’ lives easier and more enjoyable. When users feel that your recommendations truly understand their preferences and needs, they’re more likely to engage and stay loyal.

Reinforcement Learning (RL) can be a powerful primitive for Recommendation systems as it allows the system to learn and adapt to user preferences / sentiment in real-time. Enterprise adoption of RL in Recommendation systems seems to be gaining momentum with the recent availability of Cloud APIs (e.g. Azure Personalizer) and Google’s RecSim. Given a user profile and categorized recommendations, the system makes a recommendation based on popularity, interests, demographics, frequency, etc. The main novelty of these systems is that they are able to identify the features (or combination of features) of recommendations getting higher rewards for a specific user; which can then be customized for that user to provide better recommendations.

You can provide recommendations even better than Google's auto-suggest. I recently created a tool based on a prompt that seeks out search queries intended to find customers with buying intent. That's extremely valuable. Try it yourself. Imagine what you could enable your customers to search for with AI assistance.

I recently built an app, SwiftSocialAI, that merges Google Places data with the natural language abilities of ChatGPT. The app extracts information from Google Places listings to build custom social media posts. Extracting information from the listing including reviews, allows for the creation of customized posts without any user input other than the business name and location.

We did some interesting work in applying Reinforcement Learning (RL) to optimize Heating, Ventilation and Air Conditioning (HVAC) units in Factories. HVAC units are responsible for maintaining the temperature and humidity settings in a building. HVAC optimization has the potential to contribute significantly towards our sustainability goals, reducing energy consumption and CO2 emissions. Unfortunately, this is a complex problem with multiple variable factors, e.g. occupancy, manufacturing schedule, air flow dynamics, weather conditions, etc. We showed that RL is a good fit as it was able to learn and adapt to multi-parameterized system dynamics in real-time. It provided ~ 25% energy savings on top of the previously used PID controllers.

I delved into an innovative project employing reinforcement learning to train agents in a customized maze game. Using a combination of deep Q-networks and policy gradients, the agent learned not just to navigate, but also to strategize against dynamic challenges introduced in the environment. The real revelation was observing the agent's evolution, from naive explorations to mastering complex maneuvers, mimicking a human-like learning curve. It emphasized to me that the crossroads of gaming and AI is a playground, not just for entertainment but for pioneering advancements in adaptive intelligence.

In corporate education, AI has become a fundamental tool to support employee development, helping to: - Text Production and Synthesis - Construction of personalized Learning Journeys - Dimensioning of knowledge level x content offering - content production For those who have knowledge, he was a facilitator in the construction of Storytelling

While the fascination with computers playing against human players has been an obsession since before I was even born, I think I would love to see it evolve into something more akin to playing against boss levels that evolve. A lot of us struggled to get past the final dragon in Mario to save the princess, I am sure a million examples out there. It would be fun if these agents learnt from human players, adapted to our behavior. I remember working on NPCs who changed their responses based on player interaction and sentiment. It was a very nascent example of the endless possibilities of smart agents in games.

Airports and subway stations are adopting facial recognition technology for security reasons. This technology can assist law enforcement in locating individuals who are wanted or missing. Additionally, computer vision can help in identifying potential threats based on people's behavior. However, there is an ongoing and significant debate in various sectors of society regarding the potential risks associated with this technology, with a particular focus on concerns about ethnic biases.

Google is leading an initiative on FlowGPT to encourage Community Development, Education, Environmental Sustainability, Mental Health, Social Justice. The future will be profoundly impacted by AI in many incredibly positive ways. How can you get involved? Start building AI for social good and you'll change lives and probably get rich in the process.

AI holds transformative power but also poses ethical challenges. The real impact likely falls between utopian and dystopian outcomes, emphasizing the role of companies like Next AI Labs. Founded on the ethos of social good, we're committed to responsible AI use with IXCoach. We offer personalized coaching using AI, but we're equally vigilant about ethical AI practices. Our mission isn't just about business metrics; it's about making a positive, ethical impact, serving as a model for AI's potential for good while mitigating its risks.

AI plays a big role in today's world. These technologies we've talked about have many benefits, like improving security and healthcare, as well as making entertainment better. But there's an ongoing debate in society about the possible problems with AI, such as unfair treatment, privacy issues, and unintended consequences. To handle these challenges, it's important to have clear rules and guidelines for AI. They should be like guardrails that ensure AI is used fairly and safely. By sticking to these rules and keeping an eye on AI systems, we can use AI's advantages while reducing its possible downsides, making the future of AI more responsible and for the benefit of everyone.

The hardest part of building an AI project is not entirely coming up with an idea, it is what comes in between the idea and a deployed project. Here's my approach to building and testing my ideas quickly: find the best model to implement my idea, start coding the backend, and finally use Streamlit to prototype my front-end and get it out to the world! Streamlit let's you build data and ML apps quickly. It is also the fastest way to test the viability of your idea if your end goal is to find a product market fit or just for fun as well. The key is fast iteration and getting out to the masses to play with your project so you can gather feedback and improve upon your project.

For those intrigued by data science and AI projects, you don’t have to brainstorm complex ideas. Instead, focus on everyday challenges that can be addressed. Consider creating a solution that predicts and manages fluctuations in stock prices, craft AI models that provide accurate weather predictions, helping people plan their activities and businesses prepare for climate-related challenges, create apps that provide budgeting and financial planning assistance to help individuals manage their finances more effectively, build personalized recommendation systems for online platforms, suggesting movies, books, or products based on individual preferences, enhancing user experiences.

Machine learning and Electronic Design Automation: Electronic Design Automation (EDA) has been around for decades helping electronic engineers translate hardware architectures into complex multi-billion transistor semiconductor chips that power all the devices and gadgets in our hands. The challenging part of the design flow is to simultaneously optimize large design space parameters power, performance, and area (PPA) parameters of the design, and we used to spend days/ nights optimizing in the manual trial/error method. Now EDA tools offer ML-enhanced capabilities by learning from past data that can help optimize design space in a much faster, efficient, and pragmatic way for optimal PPA design. This helps productivity in a big way!

In real-world applications, people’s ideas about your AI model’s capacities can vary widely. As the developer, you are responsible for managing these expectations. Luckily, this offers great learning potential for everyone. When I was working with legal scholars to build an NLP application for court sentences, we encountered knowledge gaps in both directions. By sharing insights from our field, we began to understand each other’s methods and limitations and gradually converged to realistic estimations of the project’s legal and technical aspects. Prioritize understanding the needs and goals of your collaborators and pay special attention to bridging knowledge gaps. This can benefit both your collaboration and the final application.