Apr 2022 – Jun 2024 · 2 yrs 2 mos

• 1. Offline Personalized Rankings of titles within a carousel :
• Personaliztion stack had two parts : Big data ML pipeline for daily training and inference ,
• RankItemsAPI for ranking items within a carousel in real time.
• Pipeline computes and stores 200 M ranks for 1 M customers everyday in DDB.
• Solved Hot Partition problem by replicating the default cluster ranks across DDB shards.
• Used caffeine in-memory cache for latency improvements. Developed in-memory rate limiter to
• control rate of requests flowing to ElasticSearch.
• Designed RankItemsAPI to meet p99.9 < 100ms
• Impact: +3.29% increase in streams and +3.54% increase in streaming minutes.
• Tech Stack : Glue Jobs, Stepfunctions, S3 , DDB, ECS , Lambda
• 2. Online Personalized Ranking of Titles within a carousel (v2 model)
• Used Sagemaker endpoint for computing ranks in realtime.
• Used DAX cluster for maintaining the desired latency SLA of p99.9 < 250 ms.
• 86M customer embeddings (~30GB data) generated everyday by the pipeline.
• Impact : 18% Cost reduction , Personalization coverage of 50% from 15% , +1.30% increase in
• streaming minutes and +1.11% increase in streams.
• 3. NextUp for movies/ last episode of series (Non ML)
• Designed and implemented nextup rule for end of series / movie based on genre, studio, recency etc.
• Impact: Driving ~4% of overall miniTV streams
• 4. Manual NextUp Pinning
• Helped in pinning nextup for end of series/movie. Pinnings will always take priority over other
• recommendations by the engine. Drove streams for newly launched shows.
• Impact : +1.11% jump in minutes streamed
• 5. Personalized NextUp for movies/ last episode of series (ML)
• Impact : +2.57% jump in minutes streamed
• 6. PersonalizationService LLD Re-architecture
• Improved LLD via decoupling , abstraction and multilayering.
• Impact: Extensible design , easy to make changes and create new features in the service.
• 7. HLD for Appsflyer Integration in MiniTV App (S2S vs SDK)

Aug 2019 – Mar 2022 · 2 yrs 7 mos

• 1. Designed and implemented NextUp from ground up for Amazon MiniTV.
• Designed and implemented a Recommendation Service which exposes a next up API. Developed
• Multithreaded extensible next up rules engine which helped in ensuring p99 latency < 100ms
• Designed and implemented a Data Pipeline which keeps the ElasticSearch database up to date with the
• video catalog data.
• Impact : +8.04% Jump in streams and +5.56% jump in minutes streamed
• AWS Tech Stack used : ElasticSearch , SQS , SNS , Lambda , ECS
• 2. Dynamic Inventory Computation & Analysis Platform (DICAP) keeps inventory snapshots across warehouses in amazon. It uses this data to compute min MRP and earliest expiry across different batches of a product dynamically.
• Worked on the following optimisations in the product workflow of DICAP in order to reduce MRP-related defects on Amazon Retail.
• Out of Stock handling during bin-checks , through which we auto-routed bin-check trouble tickets to
• the next FC with maximum inventory for faster MRP validation
• Min MRP computation improvement to publish MRP from the next lot/batch when the current MRP
• entered seems egregiously low and is under the process of validation
• Fixing inventory inconsistencies by introducing atomicity between two tables which corrected a tech
• issue that kept outdated MRP/Expiry Date displayed on DetailPages for some products.
• Reducing Defects due to wrong ordering in publishing
• Impact : 42% Defects Per Million Opportunities reduction from 54 to 31.
• AWS Tech Stack used : DynamoDB , SQS , SNS , Cron Jobs

Mar 2019 – Jul 2019 · 4 mos · Bengaluru Area, India

• 1. Ramp Up Project : Developed an Android based Dictionary App (Word Sprint) and launched it in play store.
• 2. Added a dynamic configurable payment's lower and upper limit in the add funds feature in zeta android app.

Jan 2019 – Feb 2019 · 1 mo · Bengaluru Area, India

• Worked on Urban Sound Classification Model using librosa and tensorflow.

May 2018 – Jun 2018 · 1 mo · Bengaluru Area, India

• Anomaly detection in log files using Machine Learning