Jul 2023 – Jun 2024 · 11 mos · Bengaluru, Karnataka, India · On-site

• Building India-centric Foundation LLM - Krutrim.
• Engineered an Alignment Evaluation pipeline using LLM-as-a-Judge, reducing Model Evaluation time by 90+%
• Improved Alignment evaluation scores by 25% using distilled Direct Preference Optimisation (dDPO) for HHH, Political and Religious Neutrality, Conformity Bias and Identity Bias objectives.
• Utilised DeepSpeed ZeRO Stage-3 to efficiently perform DPO within resource-constrained environments.
• Extracted & processed over 10B tokens of India-centric data from PDFs & Newspaper archives, using OCR.
• Generated 100k+ data points for dDPO, spanning various alignment objectives, using Prompt Engineering on Open and Closed source LLMs - Mistral, Mixtral, GPT-4.
• Real-time Multilingual Conversion of Live-stream
• Led a team of 2 to develop a Real-Time Multilingual Voice-Cloning pipeline that consisted of Automatic Speech Recognition (ASR), Translation and Voice-cloning Text-To-Speech (TTS) modules.
• Contributed to a team effort to broadcast multilingual versions of an English live-stream for 5 Indian languages.
• In a team, integrated lip-sync technology, and streamed content to YouTube Live with a mere 5-min latency.

Jul 2022 – Jul 2023 · 1 yr · Bengaluru, Karnataka, India · On-site

• Developing ASR Models with an objective to build a Voice Assistant on a 2W.
• Designing experiments to improve ASR performance in a highly dynamic Noise environment.
• Fine-tuned SOTA models such as Whisper, wav2vec2 on custom training dataset
• Trained a U-Net Architecture based Encoder-Decoder Speech Denoiser for dynamic noise removal
• Sped up stakeholder decision making by quickly building a Voicebot PoC with ChatGPT integration
• Explored N-gram language Modelling, Text summarization, Topic modelling, sentiment analysis

Jul 2021 – Jun 2022 · 11 mos · Mumbai, Maharashtra, India · On-site

• Center for Machine Intelligence and Data Science
• Intelligent Fault Diagnostics of Rotary Machinery using Machine Learning | Project
• Objective: To diagnose multiple fault modes in rotary machines using vibration, electrical & acoustic signatures
• Gaining an insight into various fault modes using Machine Design domain knowledge
• Extensive literature survey on the use of Signal Analysis & ML for fault diagnostics
• Data acquisition
• Dataset creation for multiple fault modes by manually introducing faults/defects
• Feature Selection & Extraction
• Signal pre-processing methods such as De-noising, Compression, Time-sync averaging
• Signal processing in Wavelet, Frequency, and Time domains for feature extraction
• Fault diagnosis
• Experimentation with Machine Learning approaches such as SPC, KNNs, SVMs, HMMs

Jul 2019 – Jun 2021 · 1 yr 11 mos · Mumbai, Maharashtra, India · On-site

• Engineering Graphics Lab
• Managed smooth execution of lab sessions, examinations, and grading of around 600 students in four semesters.
• Proposed, planned, and implemented changes to methodology such as continuous evaluation for optimum work allocation

Jun 2018 – Mar 2019 · 9 mos · Pune Area, India

• Responsible for creating detailed solutions to the questions posted by students worldwide | Solutions included explanations, mathematics, charts, illustrations and references

Jun 2016 – Jul 2017 · 1 yr 1 mo · Pune Area, India

• Designing and modelling of machinery components and large steel structural units.
• Creating fabrication detailing drawings for the modeled machine components and structures.
• FEA meshing of machine components on Hypermesh.