Aug 2022 – Feb 2023 · 6 mos · Bengaluru, Karnataka, India · Remote

• Ads Monetization: Led the development and launch of Lookalike Audience Targeting feature, resulting in a 25% increase in click-through rate through A/B testing, and collaborated with cross-functional teams to bring it live on our advertising platform

Mar 2020 – Aug 2022 · 2 yrs 5 mos · Bengaluru, Karnataka, India

• 1. Offline Commands: Slot normalizer for multiple capsules(apps), fully offline/on-device solution, successfully deployed on the latest flagship mobile device(S22), for both English & Korean languages.
• 2. Contextual NLU - Was responsible for algorithmic design, development and deployment of context based partial matching for visual devices(like TV), were in direct communication with HQ counter-part.
• Used combination of tri-gram and h-trie methods to reduce the context indexing time from O(n^2) to O(n)
• Used caching to avoid repetitive computations in the parent module, reducing the run-time latency by ∼70%
• The solution was deployed for en-US, Korean and European languages after clearing all release cycles

Jan 2019 – Mar 2019 · 2 mos · Suwon, Gyeonggi-do, Korea

• Worked on Robot Skillset Generation and developed 2D Navigation module for robot using Unreal Engine.
• Used Dynamic-Window Approach(DWA) to calculate "optimal" velocity vectors for smooth navigation,
• successfully ran both simulated and hardware robot along the user-defined destinations.
• The project was highly appreciated by Samsung executives at South-Korea(HQ). Also got nominated for
• Best contributor and Team player awards for this work.

Aug 2018 – Mar 2020 · 1 yr 7 mos · Bangalore, India

• Named Entity Correction - Developed solutions using multiple fuzzy matching methods like ULA(Universal Levenshtein Automata) or “FAISS by facebook”, corrected ∼30% ASR mistakes on a 20K test data-set
• Also developed an efficient phonetic matching module using WPSM(weighted phonemic substitution matrix), which further corrected the phonetic mistakes in ASR.

Dec 2016 – Jan 2017 · 1 mo · Mumbai Area, India

• Developed a DNN based cascade framework for the reliable detection and recognition of Number plates in a vehicle
• Trained a region proposal network to identify probable regions of the number plate, followed by a CNN
• based binary classifier adapted for classifying the identified regions based on existence of number plate
• Used a Bi-LSTM network in order to generate the alphanumeric sequence written on the number plate, and achieved an overall accuracy of 91% on an in-house dataset

May 2016 – Jul 2016 · 2 mos · Mumbai Area, India

• Implemented different models for time-series prediction system including RNN and LSTM, and was able to predict whether the demand for a specific product category would increase or not with 80% accuracy.
• Developed an end-to-end pipeline for the same, from fetching data from MySQL Database to displaying
• the prediction results on an interactive GUI
• Also, proposed the system architecture for consumer-targeted product recommendation system using
• combination of content-based and collaborative filtering techniques