Mar 2025 – Present · 1 yr · Hyderabad, Telangana, India · On-site

• | Multi-Cloud Deployments | SLM/VLM Fine-Tuning | Medical NLP & Vision
• Cloud Infrastructure & Deployment (AWS, Azure, GCP):
• o Deployed Dockerized Node.js servers on AWS EC2 using PM2, and hosted React apps via S3 + CloudFront for scalable frontend delivery.
• o Ran MongoDB Atlas locally as a Docker image, enabled vector indexing, and built RAG-based search engines using embedded document stores.
• o Leveraged Azure Blob Storage for managing radiology reports and generated vector indexes with GPT embeddings using Azure OpenAI RAG workflows.
• SLM Fine-Tuning & Medical Summarization:
• o Fine-tuned models like T5, Phi-4, and Mistral-7B for radiology report summarization using transfer learning and parameter-efficient techniques like LoRA and QLoRA.
• o Used rationales from Gemini, MedLM, and GPT-4o to teach slm models using minimal datasets.
• o Evaluated models using ROUGE, BLEU, and BERTScore and served them via TensorFlow Serving for memory efficiency.
• Radiology Report Generation (Vision-Language):
• o Built pathology-specific multi-label classifiers from DICOM images using Google CXR embeddings and lightweight MLPs.
• o Developed prompt-based filters on a radiology image using cxr embeddings.
• o Dockerized all models and deployed on Google Cloud Run for scalable and stateless inference.
• Medical Data Extraction & OCR:
• o Worked on discharge summary extraction using VertexAI models like Gemini and Google GenAI.
• o Fine-tuned PaLI-Gemma by combining SigLIP-2 (vision encoder) with Gemini (language decoder) to extract structured medical data from handwritten notes using medical OCR pipelines.

Sep 2023 – Mar 2025 · 1 yr 6 mos · Hyderabad, Telangana, India · On-site

• AI Data Platform for Healthcare (ABDM Integrated):
• Contributed to an AI-powered healthcare data platform integrated with ABDM ecosystem, completing three major milestones: Patient Registration, Health Information Provider (HIP), and Health Information User (HIU).
• Built FHIR-compliant documents using AWS HealthLake, DocumentDB, and developed frontend in React (TSX).
• Seamlessly integrated platform with ABDM via asynchronous API calls and implemented Fidelius encryption for secure communication.
• Patient Demographic Search Algorithm:
• Developed a custom patient matching algorithm using Levenshtein distance for name matching, combined with filters like age difference, mobile number, ABHA ID, and demographic closeness to identify existing patient records across HIPs.
• FHIR Resource Encoding & Decryption:
• On the HIU side, implemented decryption mechanisms using Fidelius and async APIs.
• Codified health records like Discharge Summaries, OPD Consultations, Diagnostics, Prescriptions, Wellness, Immunizations, and HIEs using international standards like SNOMED, LOINC, and DICOM.
• FHIR Engine & Validation Framework:
• Built a custom FHIR engine using ABDM’s FHIR profiles, value sets, and code systems.
• Created a parser to validate health documents and used MongoDB for document storage.
• Developed migration scripts to move documents to AWS HealthLake, leveraging AWS Pipelines to handle large document sets and status-based streaming from both MongoDB and HealthLake.
• Medical NLP Pipeline:
• Converted unstructured medical text into FHIR format using cTAKES for named entity recognition and Google NLP API to extract and classify medical entities.
• Healthcare CRM & Access Control:
• Contributed to a CRM platform for healthcare organizations using AWS Cognito.
• Role-based access control (RBAC) at account, facility, and branch levels for secure, hierarchical access management.