Souvik Kar Email & Phone Number

Seattle, Wa, Us

1. Designed and led the delivery of Starfish Enrichment Platform - a streaming system that scales Minstral LLM generated product content across 2B+ items in Amazon catalog. On a high-level the design included (a) system to prioritize items for regeneration based on maximum forecasted uplift, (b) gather input data based on product scope and model requirements, (c) hosting LLM container in EKS and integrating with a proprietary prompt manager and (d) data writer protected by guardrails like text quality checker, hallucination detection, customer A/B experiments and selling partner feedback. System is scaled to stamp 4M items per day increasing average item data completeness by 4x (10 to ~38 attribute backfills). 2. Designed Catalog Enrichment Eligibility Evaluator (CE3). Atypical of traditional enrichers that typically get executed on every product listing, this system uses a combination of current KPIs (glance views, search impressions), customer feedback (review summaries) and expected forecast (based on customer A/B experiments) to detect products that is expected to gain maximum uplift with LLM based content regeneration, thereby maximizing the ROI of costly LLM invocations. This system has been used to identify 40M items across 2 marketplaces for content regeneration in Q1 2024.3. Designed and led implementation of Product Type Classification system that is used to classify all listings with new seller content in Amazon catalog, amounting to ~100M classifications per day. The breadth of work includes (a) plugging into Amazon’s new item listing pipeline, (b) optimizing model containers across different instance types to maximize TPS, GPU utilization and minimize cost per inference (AWS inferentia was winning compute choice), (c) reconciling inference results across 20 marketplaces and product variations and (d) support self-service customizable business guardrail.

Seattle, Wa, Us

Projects:1. Developing and scaling a system that auto-corrects vendor input based on a combination of historical data based learning and static business rules. Key technologies used are NLP, Hadoop, EMR and Graph-based supervised learning. 2. Currently part of a project that scrapes Amazon catalog and tries to find potential information erors like cross-attribute mismatch, image mismatch, etc. Key technologies used are Deep Learning, MXNet and NER.

Seattle, Wa, Us

Developing and scaling a system that auto-corrects vendor input based on a combination of historical data based learning and static business rules. Key technologies used are NLP, Hadoop, EMR and Graph-based supervised learning.

Sunnyvale, Ca, Us

-Worked on a project called "Edge Insights"-Worked on multi node Hadoop framework. Used Hive, Map-Reduce jobs and Oozie workflow to streamline data analysis. Used R Script to plot data of the order of 2 million points on a single world map.-Provided useful insights behind the high latency faced by users in Yahoo pages.

Kharagpur, West Bengal, In

i) Virtual Lab for Computer Architecture: Designed a single instruction CPU and a regular CPU in a simulated environment. Added several functionalities to the simulator.ii) Verification of Railway Signaling Systems: Designed an editor where user can add track segments, signal nodes and connector nodes to simulate the movement of trains. This would help in automation of railway signaling system.iii) BDD Based Logic Synthesis for countering Differential Power Attack: The function to be synthesized is presented as a BDD. This is partitioned into smaller BDDs where it is ensured all partitions have the same length. A spice netlist for the BDD is then generated by replacing each node by an NMOS transistor.