Rishikesh Ksheersagar

Rishikesh Ksheersagar

Masters in Data Science at University of Michigan
NLP Researcher at LIT Lab @ UMich
Previously - Data Science Manager at Mu Sigma Inc.

Data Science Manager

Having completed 3 years at Mu Sigma, leading teams of 5-7 data scientists, I was promoted to the position of Data Science Manager. As a manager, I was handling 2 engagements, with over 16 data scientists (in total), which had a portfolio of 1.5M USD annually. In this role, my goals were to technically support the “teams” I was leading, maintain client relationships, conduct training sessions for new joiners and assist in Business Development.

Fortune 100 Telecom Clientele

  • Me and my team analyzed the RxMER (Receiver Modulation Error Ratio) data to devise 14 different categories of network “impairments” for ~50 million modems in near real-time.
  • I led the design and development of an Unbalanced Multi-Class Classifier to classify the network Impairments using stacked XGBoost and a Sequential Deep Neural Network built on RESNET.
  • The most challenging part was to clean the data and handle imbalance. I developed the data-preprocessing pipeline consisting of Fast Fourier Transforms, Power Spectral Densities, SMOTE, padding and other tailored-methods to tackle these challenges.
  • The model was able to achieve an F-1 score of 71% and reduced the amount of customer complaints by ~30% compared YoY. THIS IS BIG!!

Healthcare - PoC

  • Alongside my main clientele, I was tasked to lead an Innovation Thread where we were developing a realistic simulation to simulate phase-3 of clinical trials.
  • We leveraged Agent Based Models to monitor the interactions between Patient Agents, Site Agents and PI (Principle Investigator) Agents. These interactions were driven using an exhaustive Bayesian Network which was trained on real world data sourced online.
  • We developed a tool which helped the clients vizualize the emergent phenomena of Patient Trial Completion and Drop-Out rates based on the initial conditions provided. This tool also had the flexibility to modify the conditions mid-simulation.
  • The simulations generated using this tool were benchmarked against the widely used methods for Clinical Trial Planning, and it achieved over 70% similarity (on average) and 79% similarity (best case) with them, which was a big win.

On the side, I conducted over 15 training sessions of Python, ML, and SQL, training over 100 new joiners per session.

I also drove 2 RFP connects with CXOs of Fortune-100 Telecom firms and added them to Mu Sigma’s portfolio.