Software Engineer & AI/ML Developer
Passionate about developing efficient and scalable AI systems, with a focus on small multimodal models that bridge vision and language.
Get to know more about my journey
I’m a software engineer specializing in AI and machine learning, with hands-on experience in developing enterprise grade intelligent systems, building and fine-tuning AI models, and creating agentic AI solutions. I’m passionate about exploring emerging technologies and actively contributing to open-source projects.
Currently working at Alcatel Lucent Enterprise
Some of my recent work
Fine-tuned BioMegatron using NVIDIA’s NeMo framework on the BC5CDR dataset for biomedical named entity recognition (NER). The model was trained to automatically identify and extract disease entities from unstructured medical text, enhancing the ability to process and analyze clinical and biomedical documents efficiently.
Created a cutting-edge project, that harness the power of Large Language Models (LLM) and Retrieval Augmented Generation (RAG) to create a smart agent capable of answering simple natural language queries. Users can pose any question to the agent for a specific URL, and the agent meticulously analyzes the URL to deliver relevant and accurate responses. This project exemplifies the advancements in AI and natural language processing, pushing the boundaries of intelligent information retrieval.
Developed a Sobel filter–based image processing system for medical X-ray chest images to enhance edge detection and boundary identification. This approach can support medical image classification by automating the detection of anatomical structures and regions of interest.
A fine-tuned variant of google/gemma-2-2b-it, trained with TRL and distilled from DeepSeek R1 to enhance reasoning during response generation. The model also supports tool invocation.
The project leverages web scraping and regular expressions (regex) to extract key information efficiently. While it may not match the sophistication of large language models with advanced NLP capabilities, it demonstrates that lightweight, low-compute solutions can effectively address real-world tasks. This project reinforced an important principle for me, not every problem requires heavy AI; sometimes, simple parsing and regex-based approaches are all you need to get meaningful results.
Some of my recent publications
Dental panoramic radiographs are widely recommended for examining abnormalities in the teeth, jaws, and surrounding structures. They play a crucial role in diagnosing various dental and jaw-related conditions, such as impacted teeth, cysts, and tumors. Additionally, they can assess bone density changes, making them valuable for the early detection of systemic diseases like osteoporosis through opportunistic screening. Osteoporosis is often called a “silent dis- ease” because bone loss occurs gradually and painlessly, making it difficult to detect until bones become weak and fracture easily. In this research, a framework was developed to automate early osteoporosis detection using a Convolutional Neural Network based deep learning approach on dental panoramic radiographs, serving as an opportunistic screening tool. The study utilized panoramic radio- graphs from 195, which were annotated into two groups by an oral radiologist: normal (C1) and osteoporotic (C2 + C3). Region of Interest was extracted using U Net. Convolutional layers were trained on this dataset for feature extraction, and features were reduced using the proposed feature influence calculation method and heuristic feature selection method and two layers of dense layer was used for classification. This model was compared with state of art deep learning image classification algorithms like inception v3, DenseNet121 and ResNet50. The proposed framework achieved exceptional performance, with an accuracy of 94%. Precision, recall, and F1 score were recorded at 0.92, 0.95, and 0.93, respectively. The proposed algorithm offers an automated approach to predict osteoporosis as an opportunistic screening tool, enabling timely intervention.
For the past few years, the air quality in Bengaluru has been hazardous. Particulate Matter 2.5 Air Quality Index (PM2.5 AQI) is the measure of air quality of the breathing air. Forecasting the PM2.5 content in the atmosphere is crucial to alert the inhabitants and also help governments in taking required actions. Deep learning architectures have proven to be efficient in predicting PM2.5. The objective of the research work intends to estimate PM2.5 in the next hour by the time series analysis of PM2.5 values in the Silk Board area of the Bengaluru region using the proposed ensemble time series algorithm. An experiment was conducted to evaluate the promising existing time series analysis models and an ensemble algorithm was designed considering best-performing models as base learners. The obtained Mean Absolute Error was 0.031 and the R2 score was 0.993.
Technologies I work with
Let's build something amazing together
I'm always open to discussing new projects, creative ideas, or opportunities to be part of your visions.