Jan 2024 – Dec 2024 · 11 mos · Bengaluru, Karnataka, India

• Skit builds digital AI agents, covering channels across Voice, SMS, Chat and Email for banking, debt collection, and other industries.
• ◦ Led a team of 5 ML engineers, assuming the responsibilities of EM & Technical Lead. Cross-function
• collaboration with Tech and Business teams to deliver product requirements, drive R&D projects to improve company’s solutions suite, build internal software tools and optimize organization wide processes.
• ◦ Led the project of replacing legacy finite-state-machines based conversational AI stack with GenAI stack for debt collection industry in USA; The new architecture improves the collection rate by 1.25x from earlier solution.
• ◦ Led the project to redesign the system architecture to support omni-channel (voice, sms, chat and email) conversation flows with single ML stack.

Aug 2023 – Jan 2024 · 5 mos

• ◦ Observe provides SaaS platform which include analyzing interactions, evaluating performance, and coaching to the contact center workforce.
• ◦ Led the project of extending the in-house ASR model trained on conformer architecture to a multi-container microservice; Boosted inference time with Triton inference serving framework and TensorRT optimization; Taking charge of functional, stress, load testing, and vertical scaling of the service to ensure SLA requirements; The service provides a 1.2x throughput speed up and costs 1/12th of the existing solution.
• ◦ Contributed as a member of the design review team which was tasked to build available, resilient, scalable and optimised software systems.
• ◦ Mentored 3 ML engineers, providing guidance on project implementation, software engineering & system architectures.

Apr 2022 – Oct 2023 · 1 yr 6 mos

• ◦ Led the migration of ML micro-services infrastructure from AWS ECS to AWS EKS to reduce costs by 30%; Conducted a knowledge session of upgraded CI/CD, Logging & Monitoring pipeline and tools for the ML Team.
• ◦ Streamlined ESPNET’s training experiments by exploiting AWS Sagemaker and Oracle Cloud tools.
• ◦ Built an internal tool to resolve customer support tickets without engineer’s intervention; Handed over the tool to support team which increased productivity of engineering team by 40-50 hours per week.
• ◦ Led the project of extending the in-house diarization algorithm (Kaldi) to a micro-service to diarize single channel call interactions; Took charge of the functional, stress, load testing, and vertical scaling of the service and deployed into production; The service provided a 1.5x throughput and 1/6th of the cost of the previous solution.
• ◦ Optimized on-premise ASR system to reduce the startup time of the service by 5 minutes; Direct improvement in latency at peak load time-durations.

Dec 2020 – Mar 2022 · 1 yr 3 mos

• ◦ Actively involved in ML platform monitoring exercises; Built a continuous evaluation framework using Apache Airflow to streamline the annotations & evaluation of ML models; this helps in consistently tracking ML models’ accuracy which are deployed in production; Maintainer of New Relic Dashboards of multiple ML services in production.
• ◦ Owner of micro-service which serves as the bridge between the backend team and ML team. The micro-service works as the routing engine to access other REST and queue based ML micro-services.
• ◦ Developed an optimized speech based ML model to filter agent-customer call interactions on the basis of fast speaking rate, articulation issue, and tone. The algorithm resulted in a F1 score of 75%. Took charge of designing, implementing and load testing of the micro-service which got deployed in production.
• ◦ Co-developed smart micro-service for resilient on-premise ASR system, in which consumers make dynamic decisions on whether to pick/consume an audio-transcription request based on the resources available at any instant.
• ◦ Introduced the usage of LLD documentation (UML class and sequence diagrams) in the SSDLC process for structured discussions and streamlined knowledge transfer.

Mar 2020 – Nov 2020 · 8 mos

• Whisper Speech Recognition: Bixby is a voice assistant indigenous to Samsung smart devices. As a part of the Automated Speech Recognition(ASR) team, built a speech-to-whisper conversion system using CycelGAN to generate synthetic whisper signals to augment training data. Research paper accepted in INTERSPEECH 2020.
• Speech-to-Speech Translation: Presently working on speech translation problems involving conversion of Korean audio signal features to English audio signal features. Using the Transformer model as the fundamental sequence-to-sequence architecture with the addition of auxiliary decoders to train on parallel tri-phone aligned data.

Jun 2018 – Feb 2020 · 1 yr 8 mos

• 1. Grapheme-to-Phoneme (G2P): I built Copy-Augmented Encoder-Decoder Bi-LSTM based architecture to achieve state-of-the-art results. Research paper accepted at ASRU 2019.
• 2. Grammatical Error Correction (GEC): Approached the GEC problem as a sequence-2-sequence task with(hypothesis, reference) as the (source, target) sentence pair. Modified Transformer architecture to handle attention from multi Encoders in a hierarchical fashion for on-device textual processing module in smartphones to assist voice and keyboard enabled services.

May 2017 – Sep 2017 · 4 mos · Gurgaon, India

• Devised a novel way of computing soiling rate in Photovoltaic plants experiencing similar climatic conditions from limited data. Authored the research abstract for submission in leading conference on renewable energy for publication.
• Engineered a forecasting module to predict hourly active power generation by a PV plant using gradient boosted trees, achieving a correlation of 0.97. Enhanced the feature set by utilizing OpenWeatherMap API.
• Automated the generation of daily visualization reports of PV Plants portfolio in Power BI by connecting MySQL server hosted on AWS EC2 instance.
• Built a GUI application using PyQt4 to facilitate the data downloading from multiple PV plants' dashboards.

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

• Studied the low-resolution electrical energy consumption data of US consumers to analyse the pattern in vacation periods.
• Devised a sliding window algorithm to predict the vacation instances of the residents using MATLAB with precision >95% and recall >70%, which got incorporated into production.
• Worked on the energy disaggregation module to compute refrigerator consumption from low resolution, raw electrical energy consumption, data which was pushed to disaggregation pipeline.

May 2015 – Jul 2015 · 2 mos · Mumbai Area, India

• Developed a polling feature for real-time meetings in Django Web Framework which was incorporated in the android app.
• Extracted artists' names from 15,000 Facebook textual posts using Stanford POSTagger after NLTK assisted pre-processing.
• Generated Artists database by drawing out information using Wikipedia API, Youtube API, and SoundCloud API.

Aug 2014 – Jun 2015 · 10 mos · Kharagpur Area, India

• Spearheaded the development of the tutorials on various fields of Computer Science to help student community in the institute.
• Conducted regular meetings to familiarize the members with Competitive Coding and Application Development.
• Collaborated with Corporate and Investment Organisations to organize 5 coding events which saw a significant number of participants.
• Planned the Calendar of the Knowledge Sessions and officiated its implementation.

Aug 2014 – Apr 2015 · 8 mos · Kharagpur Area, India

• • Presented ideas and executed the same while building the website for 12th Annual Alumni Meet.

Jul 2014 – Mar 2015 · 8 mos · IIT Kharagpur

• Leading a team of 6 members through the work culture of Global Entrepreneurial Programme in AIESEC.
• International relations management.