Pratik Kotkar
Principal Data Scientist @ bicycle.ai | Engineer | AI
Bengaluru
8+ years of work experience
Feb 20267 mosPresent
Bengaluru
Building AI agents for ML problems
Aug 20241 yr 7 mosFeb 2026
Hybrid
Built E2E ML pipeline from model design to development and deployment.
Used Airflow, Clickhouse db as a part of deployment.
Processed data for about 150M customers
Next challenge incoming is to reduce runtime responses of models for further operations
Jun 20231 yr 3 mosAug 2024
Bengaluru, Karnataka, India
Working on Customer Eligibility Score using supervised machine learning techniques to improve current manual operations scenarios and rule based systems for customer eligibility identification.
Building a customer merchant affinity engine to optimise the discounts and improve GMV and Success Rates
Apr 20232 mosMay 2023
Bengaluru, Karnataka
Studied about Machine Learning techniques in generative AI , read research papers and implementation of RAGs and LLMs , created an app for automated content generation and moderation using LLMs and performed end to end implementation of a ML model from data collection to consumption using airflow, kubeflow, docker , AWS Sagemaker etc
Oct 20226 mosMar 2023
Lending Propensity Model :
Model built to identify the high propensity customers to buy loans from the bank.
An ensemble method was used to determine the probability of customers most likely to consider buying a loan
AWS sagemaker was used for deployment of the model , ML pipeline was setup as API endpoints needed to be exposed
This model helped filtering out the customers with high propensity which resulted in 20% improved disbursement rate
Model has a Recall of 79.5% which is pretty good for the initial level of data (less samples of disbursement available).Building a user profile with the help of multiple data sources to create a scoring system to define the approval likelihood in case the user opts to buy the loan from the bank.
Marketing Source Mix Optimization :
Marketing team needed a direction to spend the budget wisely in order to maximize the output keeping the KPIs such as D7 Load, Activation etc in control
Built a model using Linear Programming method (opensolver) which gave us the right split of users to be acquired from each source such as Google,FB, Affiliate etc with CAC and budget distributed accordingly across sources.
Changes made due to this model helped us increase 5-7% of user base Week on Week and eliminate the non performing sources beforehand saving about $20k-30k /month and spend it in the appropriate sources.
Jan 202210 mosOct 2022
Bengaluru, Karnataka, India
VKYC Demand Filtering Algorithm :
Created a demand filtering algorithm for choosing the best customers for the enablement of VKYC option.
Based on onboarding data, identified customers with the highest propensity of getting a banker’s approval.
Used XGBoost for feature engineering and selection , and generated rules for segments to create a filtering criteria for customers on the go and deployed it to production.
We’ve seen an improvement from 20% to 24% within first week of deployment, upcoming weeks to show more positive results to beat the market benchmark of 60%
Nov 20201 yr 3 mosJan 2022
Bengaluru, Karnataka, India
Jul 20182 y 4 moNov 2020
Pune
Dec 20174 moApr 2018
Pune
Pratik’s AI Stack
Education
Bachelor of Engineering, Computer Science
Maharashtra Institute of Technology
20144 yrs2018
Join Pratik and 10k+ other professionals on
Tal is the app where bosses meet engineers & builders directly, without middlemen
Download Tal Boss