Apr 2024 – Sep 2025 · 1 yr 5 mos · Bengaluru, Karnataka, India · On-site

• Zero-to-one real-time agent assist product.
• LLM + tool use "agentic" application
• enabled sales folks and demos
• started with scratch with vague problem statements
• worked with PMs and users to understand requirements
• rapid prototyping + iterations -> enterprise deal

Jan 2023 – Jun 2023 · 5 mos · Remote

• Courses:
• 1. Reinforcement Learning and Decision Making
• Paused. I am unsure if I'll continue.

Jan 2022 – Feb 2024 · 2 yrs 1 mo · Bangalore Urban, Karnataka, India

• Published in NeurIPS and AACL. Served as a reviewer at ACL, EMNLP and NeurIPS.
• Productionized LLM-based voicebot for collections. Built a custom markup language to allow for function calling and token-efficient constrained generation.

Jul 2021 – Apr 2022 · 9 mos · Massachusetts, United States

• Collaborating on multiple research projects relating to Distributed Deep Learning, Multi-Agent Systems, and others with Media Lab @ MIT (Camera Culture Group).
• Published:
• AdaSplit: Adaptive Trade-offs for Resource-constrained Distributed Deep Learning.

Apr 2021 – Jan 2022 · 9 mos · Greater London, England, United Kingdom

• Built a framework for batch inference with PyTorch transformer-based NLP models with Redis priority queues. Inputs are bucketed according to length for optimal usage of GPU. This framework is now the backbone of the entire ML stack of the company.

Aug 2020 – Mar 2021 · 7 mos · Mumbai, Maharashtra, India

• Worked under Prof. Ganesh Ramakrishnan's ML Team.
• Knowledge Tracing of Students in an Online Learning Environment.

Jan 2020 – Jul 2020 · 6 mos · Gurugram, Haryana, India

• 1. Responsible for Content Feed Optimization (Ranking content).
• Scraping (Selenium, rotating proxies) of various websites.
• Large Scale Geocoding Text with in-house B+ Tree Indexed Database.
• Finetuning Text Classifier (BERT) (PyTorch)
• Pruning Trained Model for Extremely fast Inference on CPU and GPU Instances.
• Analyzing and Visualizing User Feedback (Plot.ly).
• Built Custom Language Identification Model with customized Loss function.
• The above model was trained so that Korean Text misclassified as English is penalized much more than English Text misclassified as Korean (Two-way loss weighting) to improve content recommendations.
• Trained with noisy soft labels, annotated using heuristic labelling functions and Google Translate API.
• 2. Responsible for Audio Generation (Speech):
• Building Tool for annotating and cleaning Raw audio.
• Finetuned WaveGlow (WaveNet based) for the Data collected.
• 3. Built Non-User Specific Recommendation engine with Bandits and Differential Evolution Algorithms.

Aug 2019 – Mar 2020 · 7 mos · Chhattisgarh, India

• I taught one-on-one through Skype, talking about concepts from Optimization in ML to Encoder Decoder Models and Self Attention.

Jun 2019 – Aug 2019 · 2 mos · Pune/Pimpri-Chinchwad Area

• I was responsible for engineering Machine learning solutions to Business problems and PoCs. Some of the problems that I worked on include:
• ●Assigning priorities to Customer calls using ​Audio Emotion and Speech recognition​.
• ● Building ​Face Recognition (FaceNet)​ for an E-Receptionist, formultiple use cases, such as taking attendance and greeting employees, generating visitor passes to visitors etc.. (PyTorch)
• ●Building Custom OCR system for extracting specific information from scanned documents with Yolo v3 (PyTorch)
• ●Parsing text into structured tables from PDFs using NLP tools like NLTK and SpaCy by training custom Named Entity Recognizer.
• ●Recipe2Dish - A GAN to generate an image of a Food item given its ingredients and preparation directions (text). (PyTorch)
• ●Optimizing Inference time of already deployed code in Tensorflow using Quantization and Layer Fusion.