Oct 2024 – Oct 2025 · 1 yr · Bengaluru, Karnataka, India · Hybrid

• Built Generalized Workflow Engine(GWE) low-code/no-code platform which helps in building AI powerd workflow using Agents
• Created generic Doc Parser chunking pipeline (pdf, docx, ppt) which works for complex elements like table, list-item, images

Jul 2023 – Sep 2024 · 1 yr 2 mos

• Leading 2 teams( Foundational Model training & MlOps) consisting of 10 people.
• Developed Indic tokenizer within 100K vocab with ~2 token to word ratio for all Indic Languages.
• Gathered the Indic data ~2T tokens for pre-training Indic LLMs.
• Trained a multi lingual LLMs which has capabilities to understand 10 Indic languages.
• Generated SFT data for Indic Languages and performed full & LORA fine-tuning.
• Deployed LLMs with high throughput & low latency with streaming output using vLLM & tensorRT.

Apr 2022 – Jun 2023 · 1 yr 2 mos

• Developed a pipeline for Customer support ticket resolution. It takes the audio recording of customer issues and pipeline provides the summary of the issues. Used wav2vec2 and whisper model to get the transcription and used GPT for summariation task.

Feb 2021 – Mar 2022 · 1 yr 1 mo

• Developed a pipeline for an Audio classification model for 60+ commands for the Ola Scooters. Developed a temporal convolution model for the classification of input commands with MFCC features with 92% accuracy. Converted a nonstreaming model to streaming model with proper quantization to be able to run it on Edge devices

Mar 2020 – Jan 2021 · 10 mos · Bengaluru, Karnataka, India

Mar 2019 – Feb 2020 · 11 mos · Bengaluru, Karnataka, India

• Kaldi
• ASR for Bixby and Samsung Home.
• Language models for Samsung's AI assistantBixby and Samsung Home.
• Voice Activity Detection using LSTMs and tdnn.

Jun 2017 – Feb 2019 · 1 yr 8 mos · Bengaluru, Karnataka, India

May 2016 – Jul 2016 · 2 mos · Bangalore

• Analysed the Flipkart Incident Management System (IMS) data for several trends, pattern and correlations associated with it
• and selected various data fields that are crucial for pattern recognition.
• Tried K-means algorithm but infeasible due to nature of data so finally used FP-Growth Algorithm.
• Designed user interface to select customised input parameters and visualise the results in form of bar charts

May 2015 – Jul 2015 · 2 mos · D-194 Moti Marg, Bapunagar, Jaipur, Rajasthan

• Created a dynamic form creator which helps user to create multiple boxes with adjustable size; introduce text with various
• font options and edit them; feature to relocate the same anywhere on the form filed; generate XML file for the form created.