Pranav Kompally

Experience

• 

  Delhi Public School, Nacharam

  Sept 2017 - Jan 2018

  Project Head

  1. Founded the Virtual Navigation System called DPS StreetView2. Built Custom-developed Maps for implementation of predefined Shortest routes.3. Deployed VR/AR-based solution on Google Map Platform with custom made checkpoints.

• 

  Makers Hive Innovations

  Apr 2018 - Jul 2018

  Electronics and IoT Intern

  1. Worked on Creating a simple Hyper-local Weather Station with data logging in the background to the cloud.2. Weather Forecasting using ARIMA Model and plotting heatmaps for various locations in an area3. Created a Mobile application for collecting a large amount of data when on the field.4. Developed a mobile application consisting of various icons depicting various grip patterns and performed app-based control over the Bluetooth for the same.5. Learned various nuances about work-culture of an emerging startup Show less

• 

  Makers Hive Innovations

  May 2019 - Jun 2019

  Electronics and IoT Intern

  1. Worked on implementation of Proportional Movement of Linear Actuators powered by EMG Sensors on Brunnel Hand2. Worked on Face Recognition and Tracking using the EigenFace algorithm.

• 

  ILLENIUM AI

  May 2020 - Sept 2022

  Founder

  - Built India's first and only Smart Fitness garment : Vinci X- Vinci X is a washable compression garment that monitors 20+ muscles across the upper body and provide SOTA Muscle and fitness analytics in real-time.- Recipient of grants from Govt.of India and DST amounting ~1 Million.- Raised Seed Capital to further R&D and customer validation in partnership with multiple fitness chains in India. - Orchestrated and lead the development of benchmarking EMG technology to capture muscle movement and intensity. - Responsible for the development of Smart fabric with completely washable electrodes and Signal processing technology. Show less

• 

  Indian Institute of Technology Hyderabad

  May 2020 - Jun 2022

  Prayasee and Researcher

  (a) Responsible for developing and benchmarking EMG technology from scratch. Designed SOTA algorithms to track and isolate hyperlocal muscle impulses during involuntary muscle movement.track and isolate hyperlocal muscle impulses during involuntary muscle movement.(b) Researched Latent Space Manipulation for Depth2Mesh in Diffusion models for Text-to-Scene Virtualproduction, Lip Syncing.

• 

  RAPYD.AI

  Jun 2020 - Jan 2021

  NLP Researcher

  Developed the benchmarking Nirvana Architecture, surpassing T5, DeBarta, RoBerta, and others in performance.Responsible for creating a podcast-to-summary SaaS application in collaboration with Spotify, curating a dataset of 300 million podcast episodes.Developed a state-of-the-art (SOTA) transformer-based Large Language Model for Abstractive Summarisation with pre-training techniques, achieving a ROUGE score of approximately 50, surpassing benchmarks.Managed the cleaning, curating, and maintenance of the entire podcast audio and textual dataset on Amazon RedShift CloudDB.Utilized the Wav2Speaker recognition model to identify different speakers in podcasts and separate their conversations.Implemented Conditional Generation in transformer-based architectures with Iterative Feedback Learning for more efficient human-like summary generation.Devised pre-training techniques with relative positional encoding for accurate content judgment of podcast episodes.Applied quantization techniques to reduce the model's weight by 6x.Developed and deployed a complete Podcast-to-Summary pipeline on Flask over Herok Show less

• 

  Stealth

  Jan 2023 - Aug 2023

  AI, Head and Lead

  • Ideated, Designed and Developed the complete Data Annotation Pipeline for Computer Vision and NLP based tasks.• Proficient understanding of Transformer based architectures both in Visual and Textual Context.• Developed a Large ViT (Vision Transformer) based architecture to segment various instances in overMillion images.• Added a custom developed Text and Visual Encoder on top the image segmenter for class definition-2-annotation pipeline.• Benchmarked the developed architecture with existing SOTA such as DETR, YOLO, OneFormer, Swin Transformers etc.• Orchestrated and Managed various executable scripts to represent annotations provided by AI engine into public formats such as JSON, XML and TXT.• Ideated, Designed and Programmed the complete dataset-2-annotation production level pipeline on AWS EC2 and Fargate Instances, along with distributed learning capabilities.• Built adjacent Visualisation Dashboard for annotations using tools such as Streamlit and Gradio.• Orchestrated the development of custom developed executable python scripts to improve the efficiency of dis-tributed and parallel computing abilities.• Currently, Heading the development of Large Language Models for Named Entity Recognition and TextClassification at a zero shot learning.• Responsible for the R&D of Pre-training and Contrastive learning techniques for intent recognition and classification.• Coordinated the full-stack team to build custom model serving scripts on top of PyTorch TorchServe• Ideated and Configured Model Maintenance post deployment through Docker Containers ensuring Scala- bility.Developed the ’Expander Suite’. A platform built on top of LangChain and ChromaDB for Data Visualisation and Analysis tool supporting plug and play for LLMs such as GPT, Vicuna and Llama• Developed API endpoints to automate integration onto the visualisation platform with the generated predictions. Show less

• 

  Khoury College of Computer Sciences

  Sept 2023 - May 2024

  Graduate Research Assistant

  1. Researching on Multimodal attention architecture to analyse text, video and audio data. 2. My research aims to estimate the success of a commercial by predicting the possible attention rate the ad might fetch. 3. Ideated,planned and currently executing SOTA models at the intersection of Computer Vision and advanced NLP.

• 

  Stealth Mode

  May 2024 - now

  AI Engineering Intern, LLMs

  - Llama 3, Claude 3.5 and Mistral AI in practise- Developing custom RAG infrastructure.- ChromaDB as vector store.- Postgresql in action