Low-Resource & Cross-Lingual NLP
Building and evaluating NLP systems for under-resourced language pairs such as Japanese–Nepali, where parallel data is scarce and transfer from high-resource languages breaks down in ways worth understanding.
Santosh Gaire Sharma
I build and evaluate NLP systems for low-resource languages and clinical text — and study whether our evaluation methods measure what we claim they do.
M.S. Student, Information and Computer Science · CCILAB, Doshisha University
Osaka, Japan · Permanent Resident of Japan
News
Joined CCILAB as a Research Assistant, working on cross-sensor diffusion for wearable sensor reconstruction.
Began serving as a Student Tutor at Doshisha University, mentoring incoming master's students.
Concluded my AI Engineer internship at SoranoAI after building an LLM router, an LLM-as-a-judge evaluation pipeline, and a geospatial data pipeline on Google Cloud Run.
Started NepMedJP — building the first benchmark for patient-facing Japanese–Nepali medical summarization. Manuscript targeting COLING 2027. Details
Began my M.S. in Information and Computer Science at Doshisha University, advised by Prof. Kimiaki Shirahama.
Refinetograph published in the International Journal of Information Communication Technology and Digital Convergence. Read
Research Focus
About My Research
I work on natural language processing for languages and settings the field has largely passed over. My main project, NepMedJP, builds the first benchmark for patient-facing Japanese–Nepali medical summarization — a language pair with almost no parallel data and real clinical stakes for Nepali speakers living in Japan. There I design clinician-grounded evaluation rubrics and test whether automatic metrics and LLM judges measure anything a physician would actually recognize. In parallel, at CCILAB under Prof. Kimiaki Shirahama, I develop cross-sensor diffusion models that reconstruct missing wearable signals for robust human activity recognition. The thread across both is evaluation: knowing when a model's output can be trusted. I am applying to PhD programs in NLP and machine learning.
Outside research, I love playing football, watching movies and series, and following the beautiful game — a big-time Manchester City supporter.
English · Professional / research working language
Nepali · Native
Japanese · Elementary · JLPT N4 in progress
Medical NLP & Clinical Text Generation
Patient-facing clinical text: generating summaries a patient can act on, grounded in what clinicians judge to be faithful, complete, and safe rather than in surface overlap with a reference.
Evaluation of Large Language Models
Meta-evaluation of how we measure generation quality — whether automatic metrics and LLM-as-a-judge approximate expert human judgment, and identifying where they do not and expert assessment remains irreplaceable.
Text Summarization
Abstractive and cross-lingual summarization across encoder–decoder and decoder-only architectures, with attention to factuality and to preserving meaning across a language boundary.
Multimodal & Multi-Sensor Representation Learning
Learning shared representations across sensing modalities — including diffusion-based reconstruction of missing wearable sensor signals for robust activity recognition under incomplete sensing.
Academic Output
Publications
Manuscripts in Preparation
NepMedJP: A Cross-Lingual Summarization Approach Toward Japanese–Nepali Medical Text Generation
S. Gaire Sharma, K. Shirahama
Manuscript in preparation
Target venue: COLING 2027
In Preparation
Cross-Sensor Attention for Diffusion-Based Wearable Sensor Reconstruction
S. Gaire Sharma, K. Shirahama
Manuscript in preparation
In Preparation
Selected Work
Projects
Flagship research first. Click a project to highlight the skills behind it — or filter via Skills below.
ResearchManuscript in preparation · Target: COLING 2027
NepMedJP
The first benchmark for patient-facing Japanese–Nepali medical summarization. I built NepMedJPBench — 100 parallel source–summary pairs with clinician-authored references — and a clinician-validated evaluation rubric introducing two novel cross-lingual attributes: clinical meaning preservation and medical terminology transfer. Multi-rater clinician evaluation establishes gold-standard judgments; benchmarking spans pretrained and fine-tuned mT5, Qwen3-4B (thinking and non-thinking), Qwen3.5-4B, and frontier LLMs. A meta-evaluation then tests whether automatic metrics and LLM-as-a-judge approximate clinician judgment — and where they do not.
NLPData / ML
mT5Qwen3Hugging FaceClinician EvaluationMeta-Evaluation
ResearchManuscript in preparation
Cross-Diffusion Model for HAR
A cross-sensor diffusion framework that reconstructs missing wearable IMU signals, enabling reliable Human Activity Recognition under incomplete sensing. Trained over 14 body-worn sensors from the OPPORTUNITY dataset across three realistic failure regimes — 35% whole-body-location, 40% single-sensor, and 25% random 2–3 sensor dropout — pairing a DDPM reconstruction model with a downstream C-LSTM-A classifier. Diffusion reconstruction significantly outperforms mean-fill imputation where missing sensors correlate strongly with observed ones. Currently extending to CogAge and PAMAP2.
Data / ML
PyTorchDDPMC-LSTM-AOPPORTUNITYHAR
PublishedPublished · IJICTDC 2024
Refinetograph
Integrated web application combining SRGAN for super-resolution, Convolutional Autoencoder for denoising, and Zero-Reference Deep Curve Estimation for low-light image enhancement.
Data / MLEngineering
PythonPyTorchSRGANZeroDCEFlask
ProductionLive · gharkatha.com
GharKatha
Smart rental management platform for Nepal — landlords and tenants manage buildings and units, generate Nepali-calendar invoices, track rent and electricity payments (eSewa / Khalti / bank), and handle maintenance requests in one place.
EngineeringInfra
Next.jsTypeScriptReactNode.jsPostgreSQL
Technical Proficiency
Skills
Research and data science first — engineering and infrastructure as supporting depth. Select a skill to highlight related work.
Research & NLP
Data Science & ML
Agentic AI & LLMs
Software Engineering
supportingInfrastructure & Tools
supportingSelect a skill to see related projects above.
Career Path
Experience
NepMedJP — Cross-Lingual Medical Summarization (Japanese–Nepali)
Independent Research Project, Doshisha University · Kyoto, Japan
Research2025 – Present
- Constructed NepMedJPBench, the first benchmark for patient-facing Japanese–Nepali medical summarization: 100 parallel source–summary pairs with clinician-authored reference summaries
- Developed a clinician-grounded evaluation rubric, drawing candidate attributes from the medical summarization, factuality, cross-lingual evaluation, and patient-communication literature, then validating them through a clinician survey of importance ratings and solicited criteria — introducing two novel cross-lingual attributes: clinical meaning preservation and medical terminology transfer
- Ran a multi-rater clinician evaluation across the full attribute set, establishing gold-standard human judgments for the benchmark
- Benchmarked task-specific and frontier systems to characterize architectural behavior: pretrained and fine-tuned mT5 (encoder–decoder), Qwen3-4B in thinking and non-thinking configurations (decoder-only), Qwen3.5-4B (multimodal), and frontier LLMs
- Meta-evaluated the evaluation methods themselves, testing whether automatic metrics (ROUGE-1/2/L, sacreBLEU, chrF++, BERTScore, BARTScore, medical entity-level P/R/F1) and LLM-as-a-judge approximate clinician judgment, and identifying where physician assessment remains irreplaceable
Research Assistant
CCILAB, Doshisha University · Kyoto, Japan
ResearchMay 2026 – Present
- Architected a cross-sensor diffusion framework that reconstructs missing wearable IMU signals, enabling reliable Human Activity Recognition (HAR) under incomplete sensing conditions
- Built a complete training and evaluation pipeline over 14 body-worn IMU sensors from the OPPORTUNITY dataset, with locomotion and high-level activity labels
- Designed three missing-sensor training regimes to model realistic sensor-failure patterns: 35% whole-body-location dropout, 40% single-sensor dropout, and 25% random 2–3 sensor dropout
- Trained a DDPM-based reconstruction model with a downstream C-LSTM-A activity classifier, benchmarking reconstructed signals against mean-fill imputation
- Demonstrated that diffusion-based reconstruction significantly outperforms statistical imputation when missing sensors are strongly correlated with observed ones, while identifying regimes where simple imputation remains competitive
- Extending the evaluation to the CogAge and PAMAP2 datasets in collaboration with a visiting researcher
Member
LLM-jp (NII) · Japan
ResearchPresent
- Member of LLM-jp, the National Institute of Informatics community for open Japanese large language model research and development
- Engaging with cross-organizational work on Japanese LLMs, corpora, evaluation, and open releases for academia and industry
Student Tutor
Doshisha University · Kyotanabe, Japan
TeachingApr 2026 – Present
- Mentor incoming master's students on research-software setup, Docker environments, and laboratory computing infrastructure
- Advise on research implementation, experimental design, code debugging, and academic paper writing
AI Engineer Intern
SoranoAI (Stanford-affiliated startup) · San Francisco, CA
IndustrySep 2025 – Apr 2026
- Built an LLM router using Claude Haiku 4.5 to classify incoming queries and dispatch them to appropriate tools, reducing inference cost relative to invoking a large model on every request
- Designed an LLM-as-a-judge evaluation pipeline that scores agent responses against expected outputs, enabling systematic quality monitoring and iterative improvement
- Built an end-to-end geospatial data pipeline that converts Dynamical.org datasets into GeoTIFF, generates map tiles, and renders them as interactive React Leaflet maps — deployed on Google Cloud Run with scheduled execution for automated daily refresh
Software Engineer
TeamOne Technologies · Mahalaxmi, Nepal
IndustryMay 2024 – Dec 2024
- Full-stack development for tourism-sector clients
- Built an interactive mapping application on a structured JSON schema for routes, accommodations, and trekking data
Academic Background
Education
Sep 2025 – Sep 2027 (expected)
M.S., Information and Computer Science
Graduate School of Science and Engineering (ISTC)
Doshisha University · Kyoto, Japan
GPA 4.10 / 4.50
Co-Creation Informatics Laboratory (CCILAB) · Advisor: Prof. Kimiaki Shirahama
Thesis / Research
Human activity recognition via diffusion-based missing modality imputation; Cross-lingual Japanese–Nepali medical text summarization
Key Coursework
Statistical Machine LearningPattern RecognitionResearch and EthicsNon-linear PhysicsNeuroscienceNature-Inspired ComputingPresentation and Documentation PreparationResearch and Development
2019 – 2023
B.E., Computer Engineering
Advanced College of Engineering and Management, Institute of Engineering
Tribhuvan University · Kathmandu, Nepal
First Division · 71.76% (≈ 3.6 / 4.0)
Thesis advisor: Dr. Surendra Shrestha
Thesis / Research
Refinetograph: A Machine Learning Approach Toward Image Enhancement
Key Coursework
Artificial IntelligenceDatabase SystemsSoftware EngineeringComputer Networks
Recognition
Honors & Awards
KMMF Scholarship Nomination
Nominated by the Faculty of Science and Engineering (ISTC) as one of four students, Doshisha University.
Tuition Reduction Scholarship (50%)
Awarded to privately financed international students, Doshisha University.
Government-Registered Computer Engineer
Nepal Engineering Council.
Technical Writing
Blog
Notes on AI in Japan, NLP, and systems that ship.
Japan's AI Ecosystem in 2026: Sovereign Models, Physical AI, and the Hunt for Builders
Three defining trends in Japan's AI market in 2026 — vertical sovereign AI, Physical × AI at the edge, and demand for bilingual AI builders who can ship enterprise systems.
Japan Builds the World's Robots — So Why Is Workplace AI Adoption So Slow?
Japan leads the world in industrial robotics, yet workplace AI use lags. Santosh Gaire Sharma unpacks the talent gap, aging demographics, soft regulation, and domestic LLMs reshaping Japan's AI path.
Get In Touch
Contact
Currently based in Osaka, Japan. I am open to PhD opportunities, research collaborations across Japan, and discussions about AI, NLP, data science, and healthcare. Feel free to reach out.