Apr 2023 – Oct 2025 · 2 yrs 6 mos · Bengaluru, Karnataka, India · Remote

• Founding engineer at Clazar
• Built numerous products from scratch
• Scaled the company from 0 to $X million in ARR
• Scaled the company from 0 to 250+ customers
• Owned multiple pieces of product and core platform

Jan 2023 – Mar 2023 · 2 mos · On-site

• Part of Central Platform team
• Worked on reducing backend cost per order at org level
• Led OpenSearch to Elastic (GCP) migration for the org
• Worked closely with DevOps to set up resilient multi-cloud Elasticsearch & Opensearch ecosystem
• Reduced p90 latency of serving aggregation system (served ~2k RPS) by 70% (265ms to 80ms)
• Reduced infrastructure expenses by hundreds of thousands of dollars per month.

Mar 2021 – Dec 2022 · 1 yr 9 mos · On-site

• Extensively worked on Kafka, Flink in real time ingestion pipeline that ingested 500M records a day
• Migrated search autocomplete service to Golang (first in the org) with huge (7x) performance benefits
• Worked on cost optimization initiatives to improve backend cost per order
• Implemented Search NPS widget to gather first hand user experience
• Led a team of 4 folks, revamped, and optimised autocomplete flow by improving ranking logic
• Wrote a bunch of Spark jobs to generate data required by autocomplete & intent features
• Closely interacted with product folks and contributed in PRD preparation

Apr 2020 – Mar 2021 · 11 mos

• Worked on
• Catalogue based product index for exact and near exact match intent & category retrieval without any semantic knowledge
• Lucene based data store & offline product data generation using MR jobs
• Building & scaling serving system to host catalogue based relevance module

Jan 2019 – Mar 2020 · 1 yr 2 mos

• Query categorisation model
• Trained FastText based query categorisation model and deployed it to production
• Increased training data (query x category tuples) from 0.6M to 3.5M
• Reduced model hosting hardware requirement by 68%
• Reduced p99 model serving latency from 120 ms to 10 ms
• Baseline overall Search CTR improved by 9bps
• Baseline overall Cart-Add-Buy-Now conversion improved by 10bps
• Increased query category recall by 4.8%
• Integrated the model end-to-end with serving layer
• Model training improvements
• Reduced model training time from 2 hours to < 20 minutes using multi-gpu mode
• Integrated grid search, published offline labelled metrics for multiple training iterations
• Experimented on embedding size, modelling techniques, unbiasing of user click data
• Added generation of synthetic data in data pipeline
• Published offline metrics for quality of training data
• Owned & replaced serving layer of query meta information
• Added caching, business override capability, health metrics
• Orchestrated multiple search APIs including spell, intent, redirection & domain
• Reduced latency of the service by 37%
• Obtained consistency with other semantic serving APIs
• Owned & improved query redirection service
• Increased unit test code coverage of legacy code from 16% to 71%
• Replaced two query categorisation models with single model
• Unified header contracts w.r.t multiple clients
• Identified & fixed numerous ad-hoc tasks / bugs including memory leaks, latency issue due to huge payload, charset encoding issues, system level latency issues due to worker processes, data inconsistency in model training data generation, unwanted responses being cached etc.
• Won Search Hackathon 2019
• Awarded outstanding performance for term Jan 2019 - Dec 2019

Jun 2018 – Jan 2019 · 7 mos · Bengaluru Area, India

May 2017 – Jul 2017 · 2 mos · Yokohama, Kanagawa, Japan

• Built different deep learning models using CNTK - Deep Learning Toolkit provided by Microsoft. Classified hitter, pitcher & catcher with an accuracy of 90%.
• Classified right & left hitter and right & left pitcher with an accuracy of 74% and 70% respectively.
• Improved this accuracy to 95% by using EXIF data of the images for grouping burst shot images.
• Built a Fast R-CNN model for number detection on a player’s jersey. Obtained an accuracy of 62% which was better than Microsoft OCR API.
• Classified twelve Disney characters by building a classification model. Built a portal using AJAX to create a classification medium.
• Successfully deployed the models on Microsoft Azure as WebAPI.

May 2016 – Jun 2016 · 1 mo · Aurangabad Area, India

• Implemented UI modules like Flipboard animation, LinkedIn timeline, Evernote buttons etc.
• Improved loading time of the application. Created and edited icons using Adobe Photoshop.