About
about.ipynb
AI Engineer building production LLM systems for clinical workflows. Currently at MedForce AI, working on an NHS pilot for clinical automation: agent architecture, clinical RAG, and domain-tuned models for hepatology. Focused on evaluation and reliability of medical AI. Previously shipped computer-vision and agentic-AI products at Digital-Dandelion (London) and led NLP work at Omdena. Based in Kathmandu.
Kathmandu, Nepal
Timeline
In [1]:
experience.timeline()
Out[1]:
AI Engineer
current MedForce AI London, UK · Remote
- AI engineer on a clinical-automation platform deployed in an NHS pilot. Designed the agent architecture and clinical RAG pipeline targeting clinician documentation and decision-support workflows.
- Fine-tune domain-specific LLMs (Mistral 8×22B, Llama 3, Command R family) on curated hepatology cases for medical reasoning and structured documentation; build clinical evaluation harnesses covering factuality, citation grounding, and hallucination rate.
- Cut inference cost and latency for real-time clinical use through various caching techniques (KV cache, prompt caching, semantic response cache), request batching, and quantization; built monitoring for latency, drift, and hallucination signals.
AI Engineer
Digital-Dandelion London, UK · Remote
- Built a clinical RAG system over EASL liver-disease guidelines using a multi-route retrieval architecture: simple queries served by embedding-based retrieval, complex queries routed to page-indexed advanced RAG. Reached 97% answer accuracy vs. 85% on a state-of-the-art baseline and cut end-to-end latency by 50%.
- Trained an EfficientNet image-classification system on GCP scoring 30,000+ dental websites for modernization; lifted accuracy from 89% → 95% via targeted data augmentation and an architecture switch from the prior baseline. Released the dental scraping and extraction datasets on Hugging Face.
- Built a multimodal ranking system for JLL evaluating hundreds of commercial real-estate offices via pairwise tournament comparisons across building exterior, interior, workspace, and floor-plan imagery to identify top performers.
- Fine-tuned an LLM on hundreds of past winning creative campaigns and creative-director rubrics for Page & Page (Novo Nordisk account); validated outputs on KPIs spanning relevance, creativity, and brand-fit.
Junior ML Engineer
Omdena New York, USA · Remote
- Coordinated a 10–20-person distributed team across multiple timezones building scraping infrastructure that assembled a 46,000-article Nepali news corpus, the working dataset for the program's media-representation research.
- Built Nepali-language NLP models capturing local idioms to classify how women and marginalized groups are represented in Nepali news media.
- Designed a media-diversity scoring model with explicit progress metrics surfaced to research stakeholders.
Skills
In [2]:
utshav.skills
Out[2]:
Languages
Python 
TypeScript 
SQL 
C++ AI & ML
PyTorch 
TensorFlow 
Scikit-learn 
Pandas 
NumPy LLMs RAG LangChain NLP Computer Vision Backend & Infra
FastAPI 
Docker 
Google Cloud 
Git MLOps Open models & datasets
In [3]:
utshav.huggingface
Out[3]:
Open-source models and datasets on Hugging Face. Medical multimodal models for low-resource clinical settings, plus information-extraction fine-tunes.
- MedGemma-Nepali : Gemma-3 / MedGemma 4B fine-tuned for Nepali clinical image-text reasoning. Powers MeroDaktar.
- Llama-3-70B Extractor : structured information extraction from web content. Adapter and 4-bit variants also released.
- Dental_website_scraping + Dentalweb_extraction : open datasets from the Digital-Dandelion dental-modernization project.
Education
In [4]:
utshav.education
Out[4]:
B.Sc. Computer Science & IT
Madan Bhandari Memorial College Kathmandu, Nepal