Srivatsa Srinath

A 3 day training was conducted for a technology company to help them with building production machine learning system. Over the duration of the course, the team was split up for handling the different modules from data ingestion to the visualisation using Python.

Delivered a 2 days hands-on training to experienced data science professionals in the area of deep learning using Tensorflow. The training was focused on helping the participants understand the nuances of tuning neural networks and understand the math behind it.

Delivered a 7 day training covering inferential statistics and machine learning to a group of 9 participants who were getting started with their career at a MNC providing banking and financial services across the globe. The enthusiasm and the zeal to get to the fundamentals of every topic that was discussed made the sessions very engaging, I thoroughly enjoyed sharing my views. Especially in the younger crowd, I see increasing eagerness to delve into maths, be it statistics or linear algebra;… Show more

Delivered a 7 day training covering inferential statistics and machine learning to a group of 9 participants who were getting started with their career at a MNC providing banking and financial services across the globe. The enthusiasm and the zeal to get to the fundamentals of every topic that was discussed made the sessions very engaging, I thoroughly enjoyed sharing my views. Especially in the younger crowd, I see increasing eagerness to delve into maths, be it statistics or linear algebra; this is certainly the right way forward for the upcoming generation of data scientists. Show less

4 days training to a group of 40 students in their final leg of a 6 month course on big data analytics at S P Jain School of Global Management. Thoroughly enjoyed interacting with the students and their interest in getting to know the math and statistical concepts behind the machine learning algorithms rather than simply going through exercises in Python was the highlight of the course. I thoroughly enjoyed the challenge of delivering the advanced content in a fashion that many could… Show more

4 days training to a group of 40 students in their final leg of a 6 month course on big data analytics at S P Jain School of Global Management. Thoroughly enjoyed interacting with the students and their interest in getting to know the math and statistical concepts behind the machine learning algorithms rather than simply going through exercises in Python was the highlight of the course. I thoroughly enjoyed the challenge of delivering the advanced content in a fashion that many could appreciate. At the end of the course, I am happy to have shown the students how the theory and the concepts in machine learning enables one to solve real life problems, thereby bringing out the importance of problem formulation. All the very best to the students for the upcoming placements and I hope their zeal to learn is carried forward where ever they head out. Show less

Delivered a 5 day training to a MNC dealing with Technology and Consulting, the focus was on machine learning. This time apart from building each of the concepts through practical business examples, the participants were given business case studies to solve a team and come up with solutions. They then presented to the team resulting in very engaging sessions on the last day. Loved the experience and the resulting interactions, a memorable week indeed!

4 days of intense training provided to a group of 16 experienced folks from a IT company in India. As always, I relish the challenge of introducing statistical concepts in fashion where people are able to see immediate applications into their day to day work. Once the participants started develop a conversational level of comfort with statistics, the discussions become that much more engaging.In the last day and a half, we connected the concepts in statistics to statistical learning principles… Show more

4 days of intense training provided to a group of 16 experienced folks from a IT company in India. As always, I relish the challenge of introducing statistical concepts in fashion where people are able to see immediate applications into their day to day work. Once the participants started develop a conversational level of comfort with statistics, the discussions become that much more engaging.In the last day and a half, we connected the concepts in statistics to statistical learning principles (or machine learning) and deep dived in an algorithm each for regression, classification and clustering. At the end of the training participants begun to appreciate what it really takes to solve problems in data science rather simply going through a tool based training session. Show less

Another 2 day training covering inferential statistics to a group of 25 people working in a multinational business process outsourcing and information technology services company. It was a varied group of people with various degrees of exposure to statistics; from novice in statatistics to those who had MS in Econometrics and Statistics. The topics were introduced in a case study based approach to enable the participants to develop a very practical perspective of the concepts enabling them to… Show more

Another 2 day training covering inferential statistics to a group of 25 people working in a multinational business process outsourcing and information technology services company. It was a varied group of people with various degrees of exposure to statistics; from novice in statatistics to those who had MS in Econometrics and Statistics. The topics were introduced in a case study based approach to enable the participants to develop a very practical perspective of the concepts enabling them to immediately apply them on the job. Needless to say, very interesting discussion came up enriching everybody take away from the program. Show less

An intense 2 day training on the algorithms behind machine learning and applications of the same to an analytics team at a product development MNC. Among the best of the training sessions that I have had. The audience comprised, totaling 15, comprised of experienced analytics professionals participated in intense discussion on the algorithms behind popular classification techniques like SVM, logistic regression, discriminant analysis and ensemble methods, and anomaly detection methods, Rich… Show more

An intense 2 day training on the algorithms behind machine learning and applications of the same to an analytics team at a product development MNC. Among the best of the training sessions that I have had. The audience comprised, totaling 15, comprised of experienced analytics professionals participated in intense discussion on the algorithms behind popular classification techniques like SVM, logistic regression, discriminant analysis and ensemble methods, and anomaly detection methods, Rich interactions ensued around the issues faced by a machine learning practitioner were discussed and addressed. Show less

A 2 day training covering inferential statistics to a group of 25 people working in a multinational business process outsourcing and information technology services company. The goal was to introduce the statistics with focus on the inferential statistics. The topics were introduced in a case study based approach to enable the participants to develop a very practical perspective of the concepts enabling them to immediately apply them on the job. As always, the participants were very involved… Show more

A 2 day training covering inferential statistics to a group of 25 people working in a multinational business process outsourcing and information technology services company. The goal was to introduce the statistics with focus on the inferential statistics. The topics were introduced in a case study based approach to enable the participants to develop a very practical perspective of the concepts enabling them to immediately apply them on the job. As always, the participants were very involved and interesting discussions ensued. Show less

Yet another amazing experience of teaching a group of 16 senior consultants in the financial services domain. The training was held over a period of 4 days covering inferential statistics followed by machine learning techniques used in predictive analytics. The participants were completely engaged in the training resulting in a delightful experience in the delivery of the topics. A healthy mix of technical deep dive and hands on training ensued.

Had an awesome time delivering a 3 day training on inferential statistics to a group of around 20 people making the transition from BI to the world of analytics at an MNC, based in Hyderabad, who happen to lead the world in software products. Focus was on introducing the concepts in a very applied fashion, rather just focus on the mechanics using R. Among others, we covered sampling and sampling distributions, confidence intervals, hypothesis testing, ANOVA and made in roads into predictive… Show more

Had an awesome time delivering a 3 day training on inferential statistics to a group of around 20 people making the transition from BI to the world of analytics at an MNC, based in Hyderabad, who happen to lead the world in software products. Focus was on introducing the concepts in a very applied fashion, rather just focus on the mechanics using R. Among others, we covered sampling and sampling distributions, confidence intervals, hypothesis testing, ANOVA and made in roads into predictive modeling using linear regression. Show less

A 2 day training covering inferential statistics to a group of 25 people working in a multinational business process outsourcing and information technology services company. The goal is to introduce the participants to R and then provide an in depth coverage of descriptive and inferential statistics. R programming is introduced primarily to enable learning via extensive hands on with appropriate examples.

Delivered a 32 hours training on predictive analytics (supervised and unsupervised statistical / machine learning techniques) using R to a leading ERP product company in Bangalore. The training program spanning over 4 days started with introduction to R, followed by exposition of key statistical concepts to serve as a bedrock for the subsequent coverage of predictive analytics techniques.

Provided in class hands-on training on Business Analytics with R to a class of 25 experienced professionals at a leading MNC covering R programming, key statistical concepts to build on to discussions various machine learning concepts.

Conducted “Introduction to Analytics with R”, a comprehensive introduction to the statistical concepts allied with R programming for a Mobile VAS infrastructure company, foraying into analytics to build a BI product for telecom sector.

Brent oil prices serves as a major benchmark price for purchases of oil worldwide. The goal was to come up with several models to predict the brent oil prices on a holdout dataset and pick the ones with least error component. The various models that were used were multivariate regression, ARIMA model, change point analysis based models e.t.c. Finally a linear combination of all the models was chosen to predict the oil prices with least error.

Moodle Probe - The Statistical Way - A unique statistical analysis of the open source Moodle project through extensive use of R. Focus was to go beyond descriptive statistics into inferential statistics. The concepts that we delved into for greater depths are statistical tests using hypothesis testing, confidence intervals and two sample tests. We worked on testing several of the established hypothesis for an existing FOSS project, Moodle.

Ittiam launched a SaaS based cloud-enabled Transcode, farmOTT, a logical step of marrying its core technology strength with the newer delivery models aimed at content owners and enterprises who manage their video online. A detailed study of the cost impact of the different work flow aspects was performed. Also comprehensive market research on the competitor offering and their pricing models were conducted to arrive at the pricing strategy for Ittiam.

Analyzed Hospital market in India, Apollo Hospital’s business strategy, financials and KPIs, and made recommendations for its future growth strategy – evaluated as best in the class.

UDA (Unified Design Analyzer) Tool Development – Typical implementation flows has several steps like synthesis, placement, and routing and performance verification. Given the multitude of tools a designer needs to deal with there is a need for a central platform to help them debug and analyze the designs. The central platform that I developed helps user analyze any number of tools with any number of design snapshots loaded. This has been used across multiple projects for debugging and… Show more

UDA (Unified Design Analyzer) Tool Development – Typical implementation flows has several steps like synthesis, placement, and routing and performance verification. Given the multitude of tools a designer needs to deal with there is a need for a central platform to help them debug and analyze the designs. The central platform that I developed helps user analyze any number of tools with any number of design snapshots loaded. This has been used across multiple projects for debugging and correlation analysis. Show less

Problem Statement: Given an image identify the attributes of the image, for example sleeve type, collar type and design type of a T-Shirt

Challenges:
*) lack of a readily available comprehensive labelled dataset
*) Features being visually discernible
*) Scaling these models across hundreds of categories

Solution approach:
Exploiting boot-strapping: Training neural networks necessitates a large training data to train the parameters. However labelled data generation is… Show more

Problem Statement: Given an image identify the attributes of the image, for example sleeve type, collar type and design type of a T-Shirt

Challenges:
*) lack of a readily available comprehensive labelled dataset
*) Features being visually discernible
*) Scaling these models across hundreds of categories

Solution approach:
Exploiting boot-strapping: Training neural networks necessitates a large training data to train the parameters. However labelled data generation is effort intensive, hence you build a classifier on smaller data and use it to build larger datasets. A target of 5000 images per label was achieved before extensive fine tuning of the hyper parameters were achieved.
Using a multi class multi label detection approach, a CNN based model was built on top of VGG 16 and trained for the classification task.
Experiments performed on the different types of loss function and weighing of the inputs to counter skewness in datasets. We were eventually able to hit up to 94% accuracy on the validation dataset.
This model needs to be tuned by continually by adding newer types of photo-shoot the existing model hasn’t been trained on.

Solution benefit: Extraction of attributes of through a CNN based ML approach with acceptable accuracy was achieved for key categories like T-Shirts, Dresses. Show less

Problem statement: Finding similar images through unsupervised learning techniques, while having the flexibility to search among different dimensions.

Key challenge: Lack of labelled dataset, to try approaches like triplet loss / siamese network

Solution approach: In any startup, while developing ML solutions, the cost of labelling data is not trivial. Hence we explored the path of unsupervised learning. The goal was to be able to identify the colors and the patterns across… Show more

Problem statement: Finding similar images through unsupervised learning techniques, while having the flexibility to search among different dimensions.

Key challenge: Lack of labelled dataset, to try approaches like triplet loss / siamese network

Solution approach: In any startup, while developing ML solutions, the cost of labelling data is not trivial. Hence we explored the path of unsupervised learning. The goal was to be able to identify the colors and the patterns across individually. This approach yields a color signature, similarly a hog based feature vector is generated on image patches using GMM model. These together constitute a multiple dimensional representation of the image from the perspective of color and pattern. To identify similar images a euclidean distance based metric is used.
Technologies used: Tensorflow, python (numpy)

Solution benefit: A highly scalable image similarity identification mechanism was developed. Show less

Problem Statement: How do you rank products just given the market signals from crawled data like position, price, stock availability, reviews and ratings?

Solution approach
The fundamental business premise is that no retailer would want to promote un-popular products in a sustained fashion. There will be periodic promotions, but on an average better selling products will show up. Unsupervised learning approaches similar to page ranking was explored, but results weren’t satisfactory.… Show more

Problem Statement: How do you rank products just given the market signals from crawled data like position, price, stock availability, reviews and ratings?

Solution approach
The fundamental business premise is that no retailer would want to promote un-popular products in a sustained fashion. There will be periodic promotions, but on an average better selling products will show up. Unsupervised learning approaches similar to page ranking was explored, but results weren’t satisfactory. Using a supervised learning approach, a neural network based, product comparator was developed which learnt to use the market signals to derive a logic for picking winners. Needed a creative approach to enhance the datasets to be able to train neural networks.
Technologies used: Tensorflow, python (numpy)

Solution benefit: The ranking engine developed works at the heart of the SaaS product offering. As always the case, the model continues to be tweaked with every feedback received from the customers. This granular level of metric comparator yields itself to rank groups of products, thereby yielding a cluster level ranks as well. Show less

Problem Statement: Building a ML assisted intrusion detection system.

Challenges:
*) lack of a comprehensive labelled dataset for threats
*) very low tolerance for false positives
*) key ML modeling assumptions like IID assumptions being broken

Solution approach:
The Engineering team worked on extraction flows from the IP data gathered. I worked with the networking engineers to understand the interaction between the different features and modelled a multivariate… Show more

Problem Statement: Building a ML assisted intrusion detection system.

Challenges:
*) lack of a comprehensive labelled dataset for threats
*) very low tolerance for false positives
*) key ML modeling assumptions like IID assumptions being broken

Solution approach:
The Engineering team worked on extraction flows from the IP data gathered. I worked with the networking engineers to understand the interaction between the different features and modelled a multivariate Gaussian based anomaly detection approach. This model would gather data for close to a month and then start giving anomaly alerts to the network administrator. Over time the model would continue to learn and adapt itself to the scenarios of the client.

Solution benefit: A self correcting model which minimises false alarms was developed. This approach enabled detection of anomalies beyond rule based threat detection. These models formed the basis of anomaly detection and combined with other indicators of compromise make up an effective intrusion detection system. Show less

Trained a batch of 16 Big Data and Analytics professionals, working for a corporation which is among the big 3 in the Indian IT sector, in Inferential Statistics. The goal was to deep dive into inferential statistics and also prep them for the subsequent training in machine learning. As always I approached the topics in a very applied fashion with the goal of helping the participants gain a very rigorous understanding of the underlying statistical principles. The training was well received… Show more

Trained a batch of 16 Big Data and Analytics professionals, working for a corporation which is among the big 3 in the Indian IT sector, in Inferential Statistics. The goal was to deep dive into inferential statistics and also prep them for the subsequent training in machine learning. As always I approached the topics in a very applied fashion with the goal of helping the participants gain a very rigorous understanding of the underlying statistical principles. The training was well received with complete engagement of the audience through out the duration of 4 days. Show less

Trainer for an Certified Business Analytics Professional course delivered online. The course is spread across 10 weeks with classes during the weekend. Topics covered include:
R programming, univariate statistics, hypothesis testing, linear and Logistic regression,segmentation via clustering classification methods, tree based methods and time series analysis