Oct 2025 – Oct 2025 · 0 mo · Bengaluru, Karnataka, India · On-site

• Developed and fine-tuned added context-aware medical Speech LLMs using LoRA adapters to achieve efficient, scalable domain adaptation with reduced computational overhead.

Jul 2024 – Sep 2025 · 1 yr 2 mos · Bengaluru, Karnataka, India · On-site

• 1. Drove multilingual LLM alignment strategies for Krutrim-2 via large-scale SFT, DPO, and RL- based experiments, optimizing instruction-following and reasoning across Indic and English using distributed training infrastructure.
• 2. Built high-quality Indic-English DPO and RLVR datasets, and curated specialised corpora (e.g., poetry) to guide fine-tuning of Krutrim-2’s multilingual behaviour and reduce language and identity drift on downstreaming tasks.
• 3. For Krutrim-3 pretraining, engineered a 50B+ token persona-driven data generation and translation pipeline, integrating LLMs (Gemini-Flash, GPT-4o, Krutrim-2) and transformer models (IndicTrans2, KrutrimTranslate), integrating domain classification, and tokenized batching via NLTK to overcome context constraints.
• 4. Built a geo-spatial ML pipeline using Random Forest Regressor with 10+ engineered RTO-level features (EV penetration, OLA share, locality encoding, store density) for 1400-location EV expansion across India; analyzed feature importances and mapped underperforming regions using Folium.
• 5. Designed a high-throughput multiprocessing pipeline to extract, filter deduplicate Indic text from 96 Common Crawl snapshots (90K+ WARC files each), leveraging AWS S3, Trafilatura, FastText, MinHash for NLP dataset creation.

May 2022 – Jul 2022 · 2 mos · Bengaluru, Karnataka, India · On-site

• 1. Determined 50 centres to expand operations in Bhilwara & Gwalior districts based on percentile ranking system and discovered 55 lakh+ revenue potential by identifying 1400+ prospective customers of agricultural and dairy input products.
• 2. Achieved 0.78 F1 score on a highly imbalanced dataset containing 1 lakh + farmers' dairy milk pouring data using neural networks and deploying a website with an end-to-end ML pipeline.

Jun 2021 – Aug 2021 · 2 mos · Roorkee, Uttarakhand, India · On-site

• 1. Created a Time-Series model for the drought prediction using the rainfall acquired with Indian Meteorological Department of India over 150 years.
• 2. Analyzed and pre-processed the data using several steps of wavelet decompositions and transformation techniques of Standard precipitation indexes of the seasoned data from 123 different geographical locations across North India.
• 3. Machine Learning method: Linear Regression and Artificial Neural Network were applied using distinctive models for achieving a determinant of an average of 0.6 and 0.9 of RMSE and NSE scoring methods.