30 days challenge 💪: AI applications for enterprises 🤖
I will challenge myself to show 30 applications of AI in the enterprise. The rules :
An essential part of running a company is knowing how much money you will generate in the following month, quarter, or year. With that information, you can decide how much you can spend on marketing, hiring, or infrastructure. It allows organizations to make informed decisions about their business.
Revenue forecasting is a time-series problem. We have the variable revenue (R) that changes over time (t). We have data about past revenue (we better have that), and we are interested in how R will behave in the future. In traditional time series models (AR, ARIMA, etc), we only use information about the target variable R to make predictions about the future. With modern machine-learning models such as XGBoost, we can integrate exogenous information, too, such as macroeconomic trends, inventory information, and the number of salespeople available at each point in time. These models allow us to get much more accurate results that can make or break a company.
Many companies have a catch-all email address, such as hello@mycompany.com. Those addresses are for people who want to communicate: potential employees sharing their CVs, inquiries about your companies, and service providers trying to sell you things.
If your company is not tiny, the amount of inbound you receive can be considerable.
Instead of going through the whole list of emails, you can use Natural Language Processing (NLP) to classify the topic of each email, similar to how your email provider already filters SPAM. The ML model would read the content of the email and output a label: HR, Product-Inquery, SEO-offer, etc. Then, when employees go through the long list of emails, they can only focus on those appropriately tagged.
Many companies still use paper forms to process customer claims and requests. As a result, a person needs to read those forms and input them into the company’s system. This is a laborious and error-prone task.
If the volume of forms is large enough, the company could invest in a Computer Vision system that uses OCR (optical character recognition) and domain knowledge to automatically extract all the information from a picture of the form.
If the information is too sensitive or mistakes are costly, someone can still be shown the extracted information alongside the raw image and be prompted for a confirmation.
Very simple versions of this have been used for a long time to grade multiple-choice exams in universities worldwide.
If your company has a newsletter, it might publish new issues on a broad range of topics: our own blog covers retail and manufacutring, computer vison and demand forecasting. Not all customers are interested in all the content.
By recording the interactions of every customer with all your content (what do they click? in what way do they engage with us?), you can build a represenentation of customer as a vector of features where each feature is a piece of information on how that customer interacted with your content previously. Then, you can use clustering techniques (DBSCAN, KNN, etc) to group similar users together, assign them a “profile”name, and send the same conent to all of them.
Knowing exactly how long customers spend there is a very valuable metric for companies that offer in-person services, such as supermarkets, restaurants, coffee shops, and stores. Time in Store (TS) can correlate highly to Customer Lifetime Value (CLV), average purchase per visit, and other relevant KPIs.
One way to obtain these metrics is to use an object detection model to detect each human in a store. Then, using a tracking library such as Norfair, you can map people across frames to understand who is a new customer and who has been there for a while. Without a tracking library, the system can’t know that it is the same person still sitting at the table.
An efficient way to get a system like this up and running is to piggyback from the existing CCTV (Closed-Circuit Television) system and use the already collected footage.
Similarly to understanding time in store for individuals, the occupation distribution is interested in how many (not who) people are in a given part of a building at each point in time. These metrics are indispensable for intelligent HVAC (Heating, Ventilation, and Cooling) systems: if you know that certain rooms are only occupied from Monday to Thursday because your company has flexible Fridays, you can avoid heating and cooling those rooms, save money and energy (and the planet), as a consequence.
To implement such a system, we can use an object detection model to find all people in a room, and they use simple counting techniques to track how many people are in a given area over time. If a camera covers several areas of interest, we can add additional logic in the post-processing stage and divide the field of vision into different areas of interest.
The outcome of such a model would be a time series indicating the average number of people in a room every minute, hour, or day, depending on the required frequency.
———-
Taking notes is an integral skill of knowledge workers, yet it takes time away from engaging in a discussion.
With the recent advances in speech-to-text models, it's now easier than ever to record the audio of a meeting, transcribe it automatically, and then use an ML model to generate a summary and action items for it.
The best models for this are implemented as Large Language Models (LLMs), and openAI's GPT-4 and Whisper (for audio) remain the most relevant to date (this is not a sponsored message 😉).
Many companies have created products for this particular application that automatically connect to your meetings, so you don't have to remember to take notes. Quality and availability in non-English languages vary.
—————-
Private information is intrinsic to many domains: from banking to healthcare to education, most of the data that can be collected contains private information of patients, students, and customers that should never be revealed to outsiders.
Using NLP techniques, a model can identify all the instances of private information: age, name, SSN, nationalities, etc., and replace those values with a placeholder. This is useful because we can then share or use that information for several purposes, even train ML models, without risking data leakage.
The main idea behind these models is that they first identify data: is this a date, a number, a name. Then, we can use a second model to understand what kind of data that is and remove it or obfuscate it. With current technology, we can ran this models locally, without having to expose the data online.
There used to be a time in which non-artistic folks (such as me) had the only option to add a good picture to a blog post, slide, or post on LinkedIn was to browse the web in the hope of finding a perfect patch.
Those days are gone. With advances in generative AI, particularly text-to-image models, we can now use “prompting” and models such as Midjourney and DALL-E 3 to generate astonishing images.
For example, the image attached to this post has been generated using DALLE-3 with the prompt: Photo showcasing a neopunk-inspired computer station with futuristic design elements. The computer monitor prominently displays the DALL-E 3 interface generating a captivating image. Surrounding the computer are hyper-realistic details of advanced tech gadgets and neon-colored lighting effects.
Translating tools have been around for a while: I can remember taking a picture of a menu in Czech with Google Translate back in 2016 and translating “live” to English. Very useful when ordering food abroad.
In any case, translating tools have become much more powerful, to the point in which we can use them to solve live someone speaking, as showcased by Mark Zuckberg in a demo last year (link below).
This technology can enable collaboration across distributed teams and with clients where, previously, language would have been a barrier. A friend booked an Airbnb in Uruguay and chatted with the owner in English: everything was in English. Only when she arrived at the property did they realize that the host only spoke Spanish and that AirBnB had been translating in the background without asking!
When covid hit, it was essential to know if people were following regulations and wearing masks.
Tryolabs and many other companies developed a system based on video analytics and computer vision to detect if someone was using a mask. The model could run on a very small Jetson Nano device without an internet connection, so it was particularly useful to deploy in crowded places like shopping malls and airports.
This technology can be expanded to support other use cases, such as detecting safety helmets in construction sites, helmets by bike rides, etc.
We all have used Word’s autocorrect feature at some point in time. It has been around since the 80s. That works fine for many use cases but fails to provide a more “holistic” understanding of a text.
Nowadays, we have access to tools such as Grammarly and LanguageTool that can process text and recommend changes to fix typos and make sentences more readable, formal, or personable (I used Grammarly to write this post).
Behind the scenes, these systems use NLP and LLMs to classify different phrases and offer alternative wordings.
Access to such tools for writing emails and proposals can significantly increase a team’s productivity.
Grammarly: https://www.grammarly.com/ LanguageTool: https://languagetool.org/
Following on yesterday topic of writing suggestions, a new set of tools are incresing programmer productivity everywhere. Not just chatGPT, but GitHub copilot and similar pieces of software.
One of their main features is their ability to let you write a comment such as # a function that reads all the lines from a file and prints their lenght, and have the tool automatically generate the code for you.
The most advanced copilots take in to account your existing code to offer suggetions that make sense with your existing codebase.
On day 1 of this challenge, I discussed how revenue forecasting can be a business game changer. However, what happens if you sell but your customers don’t pay you in time? That can cause severe problems as well. Furthermore, in many businesses, invoices are Net-60 (there is a 60-day window for them to pay), which creates much uncertainty around cashflows.
If you keep a record of when an invoice was created and when it was fulfilled (you can probably reconstruct that information from historical bank statements), then you can train a model that learns to predict when an invoice will be paid, and that information can be used to administer the company better. Your Administration department will be a lot happier :)
ChatGPT took the world by storm last year and has been going strong ever since. One of the most important innovations was to offer the experience of what a natural language interface could look like (in the form of good chatbots).
Across organizations, a long time is spent wrangling data and generating reports, a critical decision-making function. The role of the data analyst has risen considerably because of this need and how hard it is for non-technical folks to get the information they need on time.
New chat interfaces will allow anyone to ask questions about their company’s data (with the appropriate level of permissions, of course) without the need to wait for a data expert to generate the report for them.
Once a technology that only belonged to Sci-Fi movies, fingerprint identification has become ubiquitous, especially in today’s smartphone-populated world. You might have a newer iPhone that uses face recognition instead. Still, the core principle remains the same: we want to use technology to uniquely identify people by their physical traits, either their fingerprints or their facial characteristics.
In fingerprint identification, we extract key points from each scan (feature engineering). Then, we use pattern-matching algorithms to find a match between the recorded features and a database of the user’s features. If there is a match, we get to unlock the device!
Have you ever uploaded a product to eBay only to realize that your pictures look unprofessional? Have you ever tried to create an emoji for Slack or a sticker for WhatsApp and realized that you need to remove the background?
In the old days, you had to open Photoshop and manually edit each image to remove its background.
Nowadays, using advanced computer vision and image segmentation, we can use machine learning to automatically remove the background of an image, adjust its colors, and remove other minor flaws. A great example is remove.bg, a website that allows you to remove the background from images for free.
Below, I have attached two images: the original depicts a laptop on a table, and the edited one in which I removed the background and replaced it with a grass-like texture.
Similarly to yesterday, if you upload a product to eBay or MercadoLibre, you might be prompted to write a description about such a product. That can quickly become tedious if there are many images involved. This is similar to the alt-text in platforms such as LinkedIn.
With GenerativeAI, we can quickly upload an image and ask the model to write a caption. For some tasks, though, the old approach still works best. Define a list of attributes to be extracted from each image: how many people are there, the setting, what color palette is used, etc. Then, train an ML model for each category to recognize it, extract all attributes from that image, and then use some template (or GenAI) to write the final text.
The image below is of me presenting about MLOps in Seattle two weeks ago (it's wild how time flies!). I used an Image-to-Text model to generate the following ALT-text: a man standing in front of a projector screen. Not bad.
Text-to-image: https://huggingface.co/tasks/image-to-text Model name: Salesforce/blip-image-captioning-base
Since the pandemic started, we have all become accustomed to having virtual meetings in Zoom, Teams, and Meet. At first, most of us struggled to find decent backgrounds to use and realized that we had to shift our living room table to avoid showing the kitchen in the background.
In the past two years, video conferencing tools have added many useful features, of which blurring your background is the most useful.
Similarly to Image Retouching, these models identify a person in the image using Image Segmentation and applying a blur mask to everything else. You might have noticed that when someone else walks behind you, sometimes you get a glimpse of them without blur. Depending on the implementation, that could be because another person was detected. A big challenge in these applications is having the models run sufficiently fast.
The image below shows me playing around with the visual effects in a Google Meet call. Can you spot the differences?
You own a very expensive machine that needs to run 24/7. It has many parts that if they fail, can cause massive damage to the whole system, taking it offline. Part A is about to fail, it’s behaving weirdly. You have about 1 hour to shut the machine down, replace the part, and return everything online.
If you replace this critical part, your downtime will be no more than an hour and the cost of that single piece. If you don’t, the downtime will much longer and you could even need to replace the whole (very expensive) machine.
Predictive maintainace encapsules a familiy of machine learning models that read an assortment of signals generated by the different parts of the machine, and detect if the behaviour is anaomalous. If the behaviour is anaomalous, they flag it as such and generate a notfiication, so that an expert can decide if its worth acting upon it: checking the part, replacing the part, etc.
As more and more people return to the office, maintaining them in a decent condition becomes of the utmost importance. Unlike the personal market, where robot cleaners like iRobot are very popular, there has been a much smaller penetration in the enterprise sector.
This is partly due to the lack of reliable models with larger batteries (capable of solving larger floor plans) and less noisy (they could work overnight, but current models are too disturbing for normal office operations.
Behind these fantastic robots, there are sensors and planning algorithms. The sensors detect obstacles (but with a lot less technology than self-driving cars), whereas the planning algorithms are based on staple computer science algorithms such as SLAM (Simultaneous Localization and Mapping)
In case you haven't seen it, this video shows a cool AI-powered robot cleaning a public restroom: https://www.youtube.com/shorts/PLReiWyqAhc.
How do you know if your company has been hit by a computer virus, or an attempt to hack into your network?
Network logs are a way of registering all the events that computers and networks perform. Once can always “look at the logs”and figure out what happended.
Unfortnuately, there are thousands and thousands of logs produced each day and most are not very easy to read by a human. AI plays a key role here, as it enables us to detect logs that are öut of the ordinary”, and that an expert should review. Since we don’t know what we are looking for, we can’t train a supervised ML model for the task, but unsupervised learning is really useful.
What happens if i increase my prices? This is a quinisetial question for almost any industry. Sometimes the answer is: you will loose customers, sometimes is you won’t. Sometimes you will end up with a net profit that’s higher than before, other times you will loose a lot of money.
If you knew your customers elasticity, then you could know how they would react to a change in price and therefore, what you should do.
Solution? Use supervised ML to learn a model for customer elasticity. For this, you will need access to reliable past transaction data that details how much you sold and at which price. This sounds simple enoguh, but you would be surprised by how many companies don’t have acess to that kind of data easily.