Ideas2IT Technologies

We worked on an experimental version of the MLflow-TorchServe deployment plugin (MLflow-TorchServe 0.0.1.dev0), which would enable developers to use MLflow to deploy models built and trained in the MLflow pipeline into TorchServe without any additional effort.

To see more of our featured work, log on to https://www.ideas2it.com/work/

MLflow is a popular MLOps tool to manage a machine learning lifecycle but lacked support for Autologging parameters, metrics, and artifacts related to a PyTorch deep learning workflow. We worked closely with the Facebook AI Research (FAIR) team to add Autologging and provided support to load the native TorchScript models. We thereby enabled the saving of extra files and additional artifacts. These features were released in MLFlow version 1.12.0.

To see more featured works, visit https://www.ideas2it.com/clients. 

Problem: Our client wanted the end to end automation of flight operations.

Solution: End to end flight operations automation was developed using MEAN stack. Includes booking, planning crew, pilot reports and reports. A modern, flexible UI was implemented using Angular to capture a large set of information on Pilot reports and dynamic display of the crew schedule.

[email protected] to bring your bold new ideas to life.

Problem: Recruiters have to go through thousands of resumes to match candidates for a specific job. It is time consuming and tiring.

Solution: We built a Machine Learning model that matches the right candidates to the right job using not just their resumes, but also social data and a number of other parameters such as GitHub activity, LinkedIn recommendations, etc. The same matching algorithm is also used to find the ideal jobs for good candidates.

The applications was completely responsive across Desktops, Tablets and Mobile devices. React (Virtual DOM, Partial Rendering, One Direction Data Flow etc.) was used to mitigate the latency issues with third party data sources.

[email protected] to bring your bold new ideas to life.

Problem: An innovative finance company in the US wanted to launch a new Credit Card and Automotive Loans to low income individuals. They wanted origination and servicing applications. Rapid Go-To-Market (GTM) was of essence.

Solution:Our in-house design studio worked with client’s marketing team to design an intuitive system  for maximizing lead conversions.  We used the latest tech stack (Angular 8, Ionic) to develop a Progressive Web Application that was responsive, personalized and supported English & Spanish. A highly scalable, cloud based, microservices solution was built on a DevOps environment that aided quick iterations to market.

[email protected] to bring your bold new ideas to life.

Problem: Build a distributed supercomputer such that the user who provides the hardware for computation gets paid without risk of fraud.

Solution: We built a distributed supercomputer where any computer on the internet running Windows, Mac or Linux could serve as a node.The nodes that are used for distributed computing are clustered using the Expectation Maximization algorithm which buckets similar nodes into a cluster. Assignment of a node for doing a computational task is done using a multi-arm bandit simulation. Financial settlements are done using Nebulas, a third generation blockchain. For managing jobs, three different schedulers were considered - Kubernetes, Nomad and Docker Swarm. Finally Nomad was chosen as it requires no external services for saving its state and stores it instead on the master node. Additionally, it is a single binary that acts both as the master node and the client.

[email protected] to bring your bold new ideas to life.

Problem: Kisscam’s previous mobile app was not getting enough engagement with users. The app allowed users to capture photos and apply overlays on them to resemble Kisscam segments in stadium sports events in the US.

Solution: We felt that the app was missing out on several potential use cases and approaches that would boost their engagement with users and grow their base.We brainstormed with the client to come up with multiple use cases; ultimately narrowing it down to the following -

• Event listing and contest feature,

• Family Friendly portrayal of Kisscam 

• Voting on their favorite entries

• Event notifications

[email protected] to bring your bold new ideas to life.

Summary:Using conversational UX and human centered design for the bottom of the pyramid

Problem: Chilasa, a non-profit organization wanted a mobile app to help rural carpenters negotiate quotes and receive orders from city-based merchants. The catch? The carpenters were semi-literate and had only used messaging apps like WhatsApp.

Solution: We used Conversational UX to solve the problem. Carpenters received work enquiries as chat conversations where pricing could be negotiated, orders could be accepted or rejected, and progress updates could be shared effortlessly.

[email protected] to bring your bold new ideas to life.

Problem: A large B2B manufacturer was losing sales because the price analytics and discount approval process requires human intervention and took about seven days.

Solution: Started delivering instant quotes using AI without human intervention that reduced discounts and increased quote conversion by about 20%.  Used historical CRM data,  Decision Trees, co-occurence matrix,Prediction engine, neural nets and deep learning. 

Enterprise, Logistic regression, SVM (Support Vector Machine), XGBooster, Decision Tree, Neural nets/ deep learning, R, Pandas, Numpy, Scikit-learn, Keras, Matplotlib, Angular.js, D3.js (Dashboard),Python (ML), Node.js, Amazon RDS, Amazon Redshift, Swagger

[email protected] to bring your bold new ideas to life.

Problem: A healthcare technology provider  found that  their report generation time using Amazon RDS was averaging about 15 minutes because of the massive amounts of data(8B+ records). This was directly impacting their customer satisfaction.

Solution: We reduced the average report generation time to 10 seconds, a 90X performance improvement despite reducing the number of RDS instances by 20%. To do this, we first identified the performance bottlenecks in their current system using Appdynamics. Then we re-architected the data replication strategy. Then we optimized encryption/decryption strategy without compromising HIPAA compliance. We made several other improvements and also implemented real-time data syncing feature using Kinesis, SQS & MySQL binlog reader

[email protected] to bring your bold new ideas to life.

Problem: An automobile manufacturer wanted to track leads and conversion across manufacturer, dealer and financier at various stages through the sales.

Technical challenge: Some of the dealers were in  rural areas in the developing world and had very poor internet and mobile connectivity. 

Solution: A mobile app was developed for the dealer in React Native while the manufacturer and financier web applications were developed in React. Isolation of component rendering from API calls to  ensure fault-tolerance even in bad networks such as 2G. Multi-language support provided through Internationalization (i18n)

[email protected] to bring your bold new ideas to life.

Summary: Gamified platform that delivers extreme performance and scalability requirements using microservices architecture

Problem: Kisscam wanted to build a highly engaging and gamified platform for conducting photo contests in sports arenas and at huge events during breaks. Users can take photos & videos and quickly add one of the many frames available in the app and post to social media. People can vote and winners are selected via voting. 

Technical challenges:  Since the action happens suddenly in just a few minutes of a game the server-side needed extreme performance and scalability.

Solution: Android and iOS apps used image optimization and resolution based photo upload. Microservices are used to scale each service individually. Implemented auto-scaling in AWS ECS to scale the application without any downtime. Performance tuning was done for scaling the application to handle huge and sudden loads. We also built a cloud storage platform for KissCam users to store media files and scaled the application to support huge arenas to the tune of 50k concurrent users.

[email protected] to bring your bold new ideas to life.

Problem: There has been an explosion of retail companies built around products like Shopify. However, it is very hard to sell to these companies as there is no public data available around these companies.

Solution: The system we have built has collected over 5 million company names and 8 million contacts and adding 10000 data daily and we are looking to reach 1 billion data points which are about 30 million companies in the next 6 months. Firm-specific data from several disparate sources is collected. ETL process with NLP techniques is used to normalize the data to a unified company and person profile. A recommendation algorithm predicts and recommends the appropriate company to target. To get sales folks to use it, the UI needs to be simple, powerful, and fast. We chose to React + Flux. We introduced messaging-based interactions instead of forms where appropriate.

[email protected] to bring your bold new ideas to life.

A supply chain financing platform with complex user and company on-boarding process, trading platform to propose fund and settle transactions. On-boarding widgets were implemented using Angular.js.

[email protected] to bring your bold new ideas to life.

This tracks the exact location of service rendered for fraud prevention. We implemented a responsive UI, so that field folks can use it on a tab using Angular.js, jQuery, OpenStreetMap for mapping core data and Leaflet for mapping UI.

For mobile-friendly interactive maps, we chose Leaflet as it offers light-weight and fast. For Mapping core data we chose OpenStreetMap (OSM) because it’s free and is used by the likes of Foursquare and Evernote which is good validation.

[email protected] to bring your bold new ideas to life.

Never miss out on a striking deal, as SnapMap brings you daily deals from retail outlets, restaurants, multiplexes and more. Once you’ve made the deal and enjoyed the service, you can rate and share it with anyone. If you’re a service provider, lead users directly to your deals with SnapMap and drive sales like never before.

[email protected] to bring your bold new ideas to life.

Problem: Build a platform that allows both iOS and Android apps with minimal changes.

Solution: Power20 follows a single pattern across its 14 different applications. The application has been built from ground-up to work with data from the server without any local setting. A single codebase/package is compiled and deployed with different configurations to enable 14 applications per platform for a total of 28 applications on iOS and Android. Any new applications can be added by only including a configuration file on the server and making the configuration change on the app. Images and media explain the routines and are widely used across the applications. SDWebImage on iOS and Picassa on Android are used for efficient and intelligent media consumption to ensure the least runtime memory footprint. This, along with AF-Networking on iOS and Volley on Android, provide a robust HTTP request mechanism

[email protected] to bring your bold new ideas to life.

Problem: The client wanted the ability for the audience to stream specific camera feeds from live events so that they can watch the event from unique angles

Solution: The project had several challenges. 

• Choosing the right backend server for streaming and rendering engine for Android platform.
• Rigorous memory management techniques to ensure best possible stream quality.
• Working on bare metal android canvas code to efficiently handle the incoming video feed.
• Intelligent switching between supported protocols (RTSP, RDP and HTTP/HTTPS) based on the connection type (LAN, WAN).
• Handling network switching scenarios and high latency use cases.
• Some video feeds had four videos stitched together. Slicing of individual videos in such cases had to be done quickly and efficiently.

After researching on many of the available streaming content rendering engines, we finally fixed our implementation on Android Streaming API. The API had a plethora of features and had support for both Android SurfaceView and bare metal OpenGL rending canvas for future scaling.

We also built a robust asynchronous HTTP request component based on the Android Volley library to achieve reliable communication between EventStream REST services and the mobile endpoint.

[email protected] to bring your bold new ideas to life.

This app gives users the options to schedule their day, share daily plans, organize events like birthday and promotion parties, invite friends, conduct polls and chat within a group. They can also sync their entire list of Google, Yahoo and Group Agendas calendars and keep a tab on all future events.

[email protected] to bring your bold new ideas to life.

Gameday is a social network for sports enthusiasts which enables tournament organizers, Sports Facility Owners and Players to collaborate. We contributed ideas to break the initial chicken-egg problem that marketplace apps have.

[email protected] to bring your bold new ideas to life.

Interesting enterprise events happening all across your IT ecosystem are converted into a live stream of events organized around Topics. Real-time analytics engine to take actions on these events like intelligent routing. For UX, we leveraged the principle of 20% of actions are performed 80% of the time. We enable this 80% via messaging. Built a Slack like front-end using Angular.js and jQuery.

[email protected] to bring your bold new ideas to life.

Problem: Our client had around 200 YouTube channels that covered 3 brands. They wanted to track over 2000 such brands for which only shallow data was available. 

Solution: We crawled millions of records to fetch data on audience analytics, along with cross-linked data from G+ & Facebook. We then used Coefficient Correlation and Linear Regression and provided the required analytics for 2000+ brands, using videos that covered just 3 brands!

[email protected] to bring your bold new ideas to life.

Problem: The client wanted one single place to view Swap Analytics for Credit Default Swaps and Interest Rate Swaps. 

Solution: We extracted, translated and loaded the data from three different sources. We built custom models using Python which was then applied to the crunched data. The resulting predictions were visualized using D3.js.

[email protected] to bring your bold new ideas to life.

Problem: Many medical events are buried in the notes made by the doctors in multiple documents and many of them have only partial descriptions. As a result, the timeline of significant clinical events is not easily discoverable. 

Solution: We identified medically significant events like ‘infections’, ‘antibiotics’, ‘surgery’, ‘x-ray’, ‘lab test’ etc. We then used NLP to arrive at the occurrence time of significant events by identifying temporal expressions like ‘today’, ‘two weeks ago’ etc. extracted from the multiple documents and correlated them with the date of document creation. We then stored this data in a structured repository for easy retrieval to construct the timeline graph of the significant clinical events.

[email protected] to bring your bold new ideas to life.

Problem: Accurately predict losses due to mortgage delinquency.

Solution: A popular short term lending portal empowers Mortgage brokers to compete with banks and large retail lenders by providing Predictive Technology and Tools. Its CoreLogic model predicts pre-payments and loan defaults. We did the segmentation to increase the accuracy to predict delinquency of a loan using the Credit Bureau, customer transactions and customer behavior data.GBM model was built to predict Front End Debt to Income ratio (FEDT), which is key for predictive delinquency.  Further, we experimented with 160 variations of GBM models with Machine learning techniques using the caret package. Various visualizations were also done to compare the actual data set and the predicted data set.

[email protected] to bring your bold new ideas to life.

Problem: Medical doctors and other practitioners need to use very technical words and domain-specific vocabulary to ensure preciseness. This precision is needed so that other practitioners understand the narrative accurately.  Patients want to understand their health conditions, diagnosis and prognosis but find the clinical narratives intimidating and very difficult to understand. 

Solution: We created a dictionary of medical terms and abbreviations from existing documentation for specific domains in healthcare. Parsed the clinical narratives, and used Natural Language Processing to ensure the grammatical correctness of the simplified narrative.

[email protected] to bring your bold new ideas to life.

Problem: It is very hard for consumers to find a product at the best price from online and physical stores. Even if a retail chain is chosen, different stores price the same product differently. 

Solution: Data Science was used to track, predict and forecast prices based on various factors. We built an application in Ruby on Rails that lets shoppers pick the best prices for products from highly popular Australian stores such as Coles and Woolworths. Great user experience through React for the web and React-Native for mobile.  Designed a search algorithm to scan product catalogs using Word2Vec deep-learning neural network and identified the cheapest product by geo-location.

Further, we simplified purchase decisions by comparing prices across brands, regions & retail branches by predicting the prospective price(using Logical Regression Models) of a product in a store for the next 30 days and presented actionable insights using D3.js for Visualization.

[email protected] to bring your bold new ideas to life.

Problem: Hospitals are liable for penalties if the 30-day re-admission rate crosses a pre-specified threshold for a particular clinical condition

Solution: We analyzed the past patient history of readmissions. Then, we clustered patients based on clinical, social and behavioral factors like associated clinical conditions, age, gender, weight, lifestyle, ethnicity, economic indicators, geography, etc. With this data, we derived a model based on the training set to predict the risk of readmission for a patient. Finally, we tested the model on the testing data set, fine-tuned it for accuracy and ran it across the new patients to predict the readmission risk.

[email protected] to bring your bold new ideas to life.

Problem: A cloud security provider wanted to detect anomalies and abnormal peaks in outbound traffic using event logs. 

Solution: We set up a continuous feed of raw data logs, using AWS Kinesis from AWS Cloud-Trail. Data was then grouped by Time, Usertype and Logtype into multiple batches. From these batches, variables were generated to feed in to the Machine Learning model that finally predicted if the data entry was an anomaly using logistic regression.

[email protected] to bring your bold new ideas to life.

Problem: It is hard to detect if a social profile is genuine or fake.

Solution: The client provides a social identifier that is a free web identity credential controlled by the user. We designed and implemented an algorithm (based on research out of Stanford University) on Hadoop using PIG to evaluate the authenticity of an individual’s profile attributes using terabytes of social networking data. We also implemented various fraud detection and prevention solutions alongside viral social features to enable users to leverage their social graph to add to their assert score.

[email protected] to bring your bold new ideas to life.