We built an ML system that automatically extracts structured information from product titles. For example, from the product title ‘Apple iPhone 12 Pro, 256GB, Pacific Blue’ the system extracts the actual product name (‘Apple iPhone 12 Pro’), storage size (‘256GB’) and colour (‘Pacific Blue’). This helps in implementing better search functionalities including filtering and removing duplicate products that only have small variations in the title.

• Automatically extract information such as product name, size, colour etc. from product titles
• Classify products into different categories such as electronics, furniture, food, fashion etc.
• Cluster similar products together for de-duplication

• Less than 0.001 seconds processing time per product (at least 30,000X speedup from doing it manually)
• More than 85% accuracy after training the AI engine on around only 1500 products (adding in more data)

A nutritional platform that creates personalised meal plans for its customers based on their preferences and dietary targets was concerned if their customers might reject their recommended meal plan, thus hurting their sales and reputation. To solve this, we built an ML-based quality control system that predicts if a customer is likely to reject a potential meal plan, so that an alternative plan can be recommended instead.

• Our ML system looks at a customer’s demographics, preferences (both likes and dislikes) and their dietary requirements to determine whether they would like a potential meal plan
• If the customer is likely to reject, an alternative meal plan is proposed and run through the quality control system again
• The ML system also takes into account the customer’s history on the platform (e.g., what kind of meal plans they liked, and what they rejected). While making a decision, the system also considers other similar customers, and their behaviours and history on the platform
• The system was deployed on AWS EC2 servers

• The ML system was able to correctly flag 80% of recommended meal plans that would have been rejected by customers

In order to meet electricity demands it is important to have accurate forecasts of how much electricity will be required. Having forecasts allows companies to ensure appropriate supply of, for example, coal if they operate a coal-powered power plant. We built an AI-powered forecasting model to predict electricity demand.

• An AI engine that is able to accurately predict electricity demand for the next 14 days
• While forecasting, the AI engine takes into consideration weather conditions which are one of the most significant factors in energy demand
• The AI engine constantly re-trains itself. This ensures that it can adapt to different changes, such as, changes in population
• The AI engine was deployed via AWS Sagemaker and a REST endpoint was exposed

• Despite the fact that energy demand can vary arbitrarily, the AI system was able to achieve an error rate of less than 5%

Our client wanted to use AI to automatically extract data from invoices and receipts. The extracted data can then be programmatically added to an accounting software such as Quickbooks. This eliminates the need for manual data entry - which is significant for larger companies that may need to process hundreds of invoices and receipts on a recurring basis.

• Automatically extract information such as billing address, due date, invoice date and line items
• Provide the location of each extracted information (e.g., where the billing address is written) to allow users to confirm that the information is correct. This helps users avoid the need of scanning through the entire invoice to find the relevant information

• Near-perfect accuracy for most invoices and receipts, including scanned ones
• End-to-end automation of the entire workflow of extracting information from invoices and receipts and adding them to an accounting software

ASOne is an open-source library that makes object detection and tracking accessible for everyone, including high-school students and educators, inspiring them towards AI-oriented careers. ASOne enables users to set up and use different object detectors and trackers in fewer than 20 lines of code and integrate it with various external devices such as IP cameras.

Link to the repository is: https://github.com/augmentedstartups/AS-One

• ASOne supports multiple versions of YOLO including v5, v6, v7 and v8 and also different variants such as tiny and large.
• Several object trackers including DeepSort, ByteTrack and NorFair
• All models have been optimised for performance and benchmarked
• ASOne can be installed via pip

• Any type of object detector can be integrated with any object tracker in under 20 lines of code
• A wide variety of models to choose from depending on hardware and compute limitations

One of our clients identified the need of an AI tool that can convert horizontal videos (which are suited to YouTube) to vertical videos (which are preferred on TikTok). This would allow content creators to re-purpose the same video for different platforms. We build an AI tool that identifies the main protagonist in each frame of a horizontal video and then crops a portion of the frame with the protagonist in the centre. This was essential because just cropping from the centre of the frame usually ends up removing focus from the main characters.

• Automatically detect the main protagonist in each frame (the AI engine looks at preceding and subsequent frames to make a decision)
• Detect different scenes in a long video so that it can be chopped up into smaller parts. This is useful for platforms where there is a time limit on uploaded videos

• Instantly convert horizontal videos with aspect ratio 16:9 to vertical videos with aspect ratio 9:16
• Extract different scenes from several-hours long videos in only a few seconds (as opposed to doing manually which takes significantly longer)

Often instead of watching an entire video, people want a summary of the main discussion that they can scan through quickly. This is especially useful for videos in which a speaker is presenting powerpoint slides. We build an AI-powered video summarisation tool that automatically extracts powerpoint slides being presented.

• In each frame of the video, identify the region corresponding to the powerpoint slide and crop that portion out
• Detect content (text and images) on each extracted slide and use that to deduplicate slides across frames
• Combine all extracted slides into a PDF

• Summarise a one-hour video in less than 2 minutes
• Near-perfect recall of identifying all slides in the video

The company's customers want personalised exercise recommendations based on their history and profile information, such as age, gender, type of exercise they do, and location. Based on an individual’s preferences and history, the system will recommend at least 5 exercises to the customer. If the customer doesn’t like the exercise, they can reject it, and the system will again recommend another exercise.

• Our ML system recommends the exercise to an individual’s preferences and history. For example, If the customer enters information such as age, gender, exercise location, whether inside or outside the home, and previous choices.
• The system uses different types of recommender system filtering, named as collaborative filtering, content-based filtering, knowledge-based filtering, and demographic filtering.
• LSTM model is used to recommend a relevant exercise to the customer.

• The recommender system is approximately 88% accurate upon giving close recommendations to an individual customer.

The shopping mall wants to detect the age of every person in the mall. For security and surveillance purposes, we designed the AI system for the mall, which detects every person and categorises on the basis of their age.

• CNN model was used for custom age detector.
• Every person’s age was categorized into different age groups such as from {0 - 5 age group}, {15 - 30 age group}, {30 - 50 age group}, and {50 - 90 age group}.
• After age detection, for example, a person with the age of 5 lies in the first group, and the same method applies for further catheterization as well.

• The custom age detection system shows 85% accurate results, and after showing those results, categorises customers for different purposes.

The company goal was to generate Kazakh language-based articles with the help of AI system. So, to generate such articles we built an AI system based on the two NLP models named as Translational model and the Language generation model.

• Our AI system will take Kazakh language text as input and then pass it from the processing it will generate text in limited words.
• The pipeline worked as; Kazakh text as an input to the translational mode named facebook/nllb-200-3.3B and converted the text into English language, then this information passes through the text generation model named OpenAI and generates the text; furthermore, this generated text again translated into Kazakh text as an output.

• After comparison between bigscience/bloomz-7b1 and text-davinci-003, we come to the outcome that the text-davinci-003 model generates more accurate results than the bigscience/bloomz-7b1.

Wildlife conservation activities often involve flying drones over groups/herds of different animals and counting the number and types of animals in these herds. An organisation working in African forests wanted to automate this entire process. To automate this, the system was divided into two solution stages. These two stages with AI system will be defined below:

• First stage - The drone’s thermal camera will capture images and detect whether animals are in its field of view.
• Second stage - All the detected images will be transferred to the server where state-of-the-art deep learning-based object detection techniques are used to count the number of animals to each species observed.

• Approximately 77% of average precision evaluation metrics were calculated after AI system training.