I build applied ML systems and measure them: retrieval pipelines with recall@k evaluation, segmentation models judged on held-out mIoU, and NLP extractors scored against labelled test sets. I care about finding where a model fails, not just showing where it works.
- B.E. in Artificial Intelligence & Data Science, University of Mysore School of Engineering (2022–2026), CGPA 9
- AI Development Intern at Kinetrix Technologies (Oct 2025 – May 2026): built Hugging Face-based clinical speech-to-text (Whisper, BERT) and radiology image analysis (YOLO, ResNet) modules for CARE, an open-source healthcare platform, behind a FastAPI backend with OAuth2/JWT/RBAC and Celery workers
- Based in Mysuru, Karnataka, India
| Project | What it is | Result |
|---|---|---|
| Lumina RAG | Private, local document Q&A with citations. FastAPI, FAISS, local LLMs via Ollama, OCR for scanned PDFs, OpenAI-compatible API, Docker | Built a retrieval evaluation (recall@k, MRR). A coverage test found the chunker silently dropped about 13% of each document; fixed and covered by tests |
| Disaster Segmentation | Pixel-level flood damage maps from drone images (FloodNet, 10 classes). U-Net with a ResNet34 encoder in PyTorch | 70.7% mean IoU on 448 held-out test images. Small objects (vehicles, pools) are the main error, analysed per class |
| Clinical NLP Demo | Turns synthetic doctor dictation into structured sections and a medication list, with negation handling and a summary checked for invented numbers | Micro F1 0.71 → 0.97 on 20 held-out labelled notes; negated findings listed as complaints 6 of 9 → 0 of 9 |
| CafeCritic | Cafe recommender: TF-IDF similarity on cafe profiles combined with rating, in Streamlit | A data audit showed every reviewer had one rating, so collaborative filtering could not work; redesigned around what the data supports |
Every project has a README with setup steps, results, what failed and what I changed, plus tests that run on each push.
ML and data: Python, PyTorch, scikit-learn, Hugging Face Transformers, pandas, NumPy, OpenCV, SQL
LLMs and retrieval: RAG, embeddings, FAISS, chunking, retrieval evaluation (recall@k, MRR), prompt design, Ollama
Serving and tools: FastAPI, Celery, Docker, Streamlit, PostgreSQL, Git, Linux, GitHub Actions