smallstep.ai

Founder (Dec 2023 - Present)

Building Marathi LLM:

• Pretraining and Finetuning of Marathi Llama - Misal 7Bn, 1Bn parameter models.
• End to end development, evaluation and deployment of Marathi Llama.
• Open sourced Instruction tuned model here
• Published a technical blog with entire procedure here
• Launch : Linkedin Post, Twitter Post
• Coverage : YourStory, Moneycontrol, ARB Podcast
• Open Source : Huggingface Space

Medpiper

AI Consultant (Jan 2024 - Jun 2024)

Medical Documents Digitization:

• Led the creation of AI platform dedicated for extraction and digitization of health records.
• Built robust solutions to efficiently handle medical documents in diverse formats across vendors.
• Developed high-performance document extraction system, processing batches with sub-second latency.
• Optimized systems to reduce processing time from 10 minutes to 1 minute, achieving a 10x performance boost.

Tekion

Data Scientist (Aug 2023 - Jul 2024)

Document AI - Table Extraction:

• Developed a robust solution for extracting rows and columns from tables in images
• Trained a mask-RCNN based model to identify the structure of table components
• Reconstruction of table components for consumption

Automobile Service Recommendation:

• Conducted comprehensive EDA on seasonal trends in services taken
• Discovered key associations between different services taken
• Generated recommendations for individual vehicles based on their unique service histories.
• Impact : Significant increase in service add to cart and service taken rates

Pratilipi

Data Scientist (Mar 2021 - Aug 2022)

Modelling user item interactions using autoencoders:

• Leveraging bottle neck vectors as embeddings
• Creating a querying model to fetch similar interactions
• Built a collaborative filtering model
• Impact : High impact in reads and improvements in monetization observed

Creating hooks for increasing interaction at the end of each content:

• Leveraging Author similarity and common interactions of users
• Generating relevant author embeddings to capture category information
• Impact : 2x growth in reads after a user completed a read

Next authors to follow model:

• Applied clustering and probabilistic approach to compute “follow author” recommendations
• Impact : 20% increase in follow action from author profile page

Data Science Intern (Dec 2020 - Feb 2021)

Category personalisation “For you” section:

• Captured category interests and subcategory interests of users
• Personalised various category combinations
• Impact : Multi front impact on top line, monetisation and author follows observed

Conducted multiple experiments in “For you” section of the app:

• Tested multiple hypothesis to increase reads
• Conducted analysis to get insights of user behaviour across multiple geographies