HARSH KHANDELWAL
ML Researcher · NLP & Generative AI Specialist
Focused on fine-tuning Large Language Models, knowledge distillation, and applied AI for the clinical domain. Building efficient, real-world medical AI systems.
01 ABOUT
ML Researcher focused on NLP and Generative AI. Experienced in fine-tuning Large Language Models (LLMs), knowledge distillation, and applied AI for the clinical domain. Passionate about building efficient, real-world medical AI systems. History enthusiast and enjoy hiking and traveling.
Friedrich Alexander University Erlangen Nuremberg
Master's of Science in Autonomy Technologies
Expected 2026 · GPA: 2.0
National Institute Of Technology Delhi
Bachelor's of Technology in Electrical and Electronics Engineering
2020
02 EXPERIENCE
Data Science Working Student
Siemens Energy – Nuremberg
- Automated Snowflake ETL pipeline ingesting SAP data, enabling real-time Tableau dashboards
- Reduced manual data retrieval by 1 hour per analysis for project managers
- Created Power BI visualization dashboards, streamlining team workflows
AI Research Volunteer
ScaDS.AI Lab, TU Dresden – Dresden
- Contributed to research on bilingual language conditioning in LLMs published at IJCNLP-AACL 2025
- Analyzed model interpretability in Llama3-70B and Qwen-72B
- Developing universal pipeline for evidence-based retrieval-generation system
AI Workflow Intern
National Instruments, Emerson – Dresden
- Awarded performance bonus for 2 invention disclosures submitted to patent committee
- Developed manufacturing debug assistant with GenAI and tool-calling agents
- Achieved 80%+ accuracy on 90% of test failures
- Implemented data drift detection pipeline using statistical tests
- Reduced manual analysis time by 8 hours per report
Student Assistant ML/Radar
Inst. of Microwaves and Photonics, FAU – Erlangen
- Developed ML-enabled radar-based human pose estimation system
- Conducted infrared and radar motion capture studies
- Optimized RNN inference speed by evaluating lightweight architectures
Consultant, Associate Consultant
Adnate IT Solutions Ltd – Jaipur
03 PROJECTS
RAG Pipeline for XNAT-based PACS
Developed a RAG prototype for an XNAT-based PACS system using a sample dataset, targeting deployment at University Klinikum Erlangen. Processed DICOM and NIfTI CT files to populate a PostgreSQL database with pgvector, enabling vector similarity search on image embeddings.
Dockerizing Medical Report Model
Created a docker container of a VLM model for experimentation at the University Klinikum Erlangen for CT scan report generation. Built a pre-processing pipeline to convert CT-Scans to model requirements and a translate layer for English to German reports.
Finetuning GRITLM for Evidence Based Retrieval
Finetuned GRITLM embedding model for evidence based retrieval generation. Created datasets for finetuning GRITLM for the CheckThat!2025 challenge in preparation for CheckThat!2026 at CLEF 2026.
MobileLLM-Distill: On-Device LLM
Implemented 1B-parameter student model from scratch with knowledge distillation from 3B teacher (LLaMA 3.2), achieving desktop inference at 45 tokens/sec and mobile-ready INT8 quantization under 1 GB. Built end-to-end reproducible pipeline for mobile deployment targeting Snapdragon 8 Elite SoCs.
Content-Based Podcast Recommendation
Reduced data processing time by 50% with Python pipeline processing 10K+ RSS feeds using parallel operations. Developed semantic recommendation system achieving 85% accuracy by generating sentence-transformer embeddings from episode metadata and implementing cosine similarity matching.
04 CONTACT
Ready to collaborate on cutting-edge AI projects? Let's connect and build the future together.