AI Development Services

Background:

The company provides project flow management software to enhance on-time project delivery. Gantt is a web-based multiuser lanning interface, but its complex architecture poses usability challenges, even for experienced users. The company aimed to reduce reliance on videos and PDFs by introducing a chatbot as a guide.

Challenges:

• Extracting information from videos and PDFs
• Enabling the chatbot to respond to pertinent questionsfrom the video within 1.5 seconds and summarize the content of the PDF within 9 seconds, all without using a streaming API
• Ensuring consistent answers for semantically similar questions

Solution:

• Extracted transcripts from videos and combined them with PDFs to create a knowledge base
• Integrated the knowledge base with the ChatGPT model to answer questions
• Implemented semantic caching and asynchronous generation of answers

Results:

• Achieved a 90% accuracy rate in summarizing answers
• Provided correct video URLs over 80% of the time

Background:

We teamed up with an animation company. They wanted us to use GenAI to create audio software that could mimic artists’ voices. This way, they could fill in missing parts during dubbing without having the artist physically present.

Challenges:

• Extracting the artist’s voice with emotions from audio samples
• Removing noise from training data for clear sound

Solution:

• Trained the model using a 2-minute video with the artist’s voice on a T4 GPU
• Captured artist’s voice nuances using advanced machine learning techniques
• Implemented the solution using Retrieval-based-Voice-Conversion architecture

Results:

• Predicted a realistic and clear voice with source voice emotions and words
• Successfully mimicked the artist’s voice, eliminating the need for manual recording
• Demonstrated rapid convergence, highlighting its efficient learning process and ability to replicate the target voice within a short timeframe.
• Got 0.9 speaker accuracy (measured using cosine similarity of SV model embeddings)

Background:

A company in the email marketing domain aimed to boost email engagement using AI-generated images and AI subject line helpers.

Challenges:

• Ensuring captivating content and images for increased open rates
• Ensuring authors avoid wasting time due to challenges presented by prompt engineering
• Integrating AI solutions into the existing email creation tool
• Managing communication between email content and AI-based solutions

Solution:

• Implemented an AI-based solution for text and image content creation
• Integrated a ChatGPT-based solution that offers AI cards like ‘Proofread,’ ‘Condense,’ ‘Rewrite,’ and more for text-based content creation
• Adopted a two-way communication approach for image generation. Initially, ChatGPT assists in crafting a precise AI image search prompt. This prompt is then utilized to search for images using DALL-E, streamlining the process and reducing the iterations needed to find the perfect image
• Introduced an AI-based subject line helper using ChatGPT. This feature generates subjects based on email content, aiming to enhance the email’s open rate

Results:

• Increased email open rates by 7% with engaging subject lines
• High-resolution images resulted in more user responses
• Received great reviews from existing customers, and new customers have also shown interest in the product

Recognizing the shape of cells and detecting cells under mitosis is a challenging problem. Classical image processing was used for cell localization, image transformation and feature extraction. A trained convolutional neural network was used to classify the cell into different classes. However, to detect similar regions in a H&E stained tissue sample unsupervised approaches were used. Here a neural auto-encoder played the role of an image fingerprinting system.

Devices that capture multiple electro-mechanical features generate huge amounts of time-series data. We determined which set of features were significantly contributing to mechanical failure in different parts of the machine. Moving windows of time-series statistics were used to model conditions that represent the current state of the machine leading up to a future breakdown.

Processing text inputs can be challenging. NLP algorithms needed to work for multiple languages, cope with the use of slang and be able to process short-form text that has partial grammatical consistency. We used a combination of techniques involving parse tree traversals, a creation of overlapping n-gram sequences, language detection APIs and auto-encoder based pattern matching engine to extract concepts and questions from chat conversations or short reviews. These techniques can work for small data without the need for large trained models. Sequence2Sequence techniques were used to automatically convert a block of text into key-value pairs of attributes and their values. The final solution was an API-based service that was integrated into the main product.

Factors that influence rental or sale value of homes depends not only on the structural parameters of the home but also on geographical parameters and chronological factors. Multiple models were created after feature analysis while accounting for partial data and in some cases very little data for a region. Boosting based algorithms for regression were employed and the core implementation was adapted for a case where a price range needs to be predicted instead of a point estimate. Thousands of models were created and deployed so that models suitable for a region can be used to make the predictions. The final solution was an end-to-end flow adapted for the product architecture.

Higher prediction accuracy was achieved using an ensemble consisting of a trained deep neural network, a pre-existing application-specific predictor, and a support vector machine classifier. While the DNN acted as a new complementary predictor, the SVM served to determine which of the two predictors probabilistically gave the correct result at that moment. This is achieved by creating a probability surface that exhibits a bi-model distribution peaking at each of the predictors. Such techniques can be used in applications such as indoor locationing, object location predictions in images, voting based systems

In a multiparty contest involving rating by human evaluators, we algorithmically determined which of the many contest ideas should be presented to the human evaluator such that a likelihood estimate is maximized. In another situation, the objective function of a system of non-linear equations was minimized with constraints on the parameters. Here we solved the problem using Lagrangian relaxation with gradient descent to determine the optimal parameters.