Graduate Employee

September 2026 - Present

University of Oregon

Eugene, OR, USA

• Lead data science labs (DSCI 101) for approximately 30 students, guiding exercises and clarifying data science concepts.
• Hold office hours to address course questions and provide feedback on quizzes.

Software Engineer I, Machine Learning

November 2025 - August 2026

Smart Data Solutions

Nepal

• Model training and extraction: Fine-tuned vision-language models with full-parameter supervised training on approximately 40,000 healthcare documents, preparing the data and field-specific prompts. Unanimous field-level consensus across three models improved overall extraction accuracy by 21 percentage points over the existing system.
• Batch inference: Implemented vLLM batch inference for fine-tuned models, contributing to an observed approximately 60% reduction in per-document latency. Collaborated with the systems team on deployment.
• Multimodal RAG: Built structured PDF extraction combining dense/BM25 retrieval over OCR text with a separate ColPali visual retrieval branch.

Machine Learning Engineer

January 2025 - October 2025

Root Level AI

Kathmandu, Nepal

• Retrieval evaluation: Built benchmarks with manually authored queries to compare embedding models and dense, sparse, and hybrid retrieval. Evaluated retrieval performance separately from generated-answer quality using retrieval metrics and answer-similarity scores.
• Conversational assistant: Built function-calling workflows with conversation memory, a hybrid retrieval tool using Reciprocal Rank Fusion (RRF), and human handoff controlled by configurable escalation instructions.
• Document ingestion: Built Amazon S3 and AWS Lambda components for batched PDF/OCR processing with bounded concurrency and retries, supporting multitenant Qdrant retrieval.

Machine Learning Engineer

April 2024 - January 2025

DoriIT

Kathmandu, Nepal

• LLM integration: Built RAG assistants with OpenAI and Gemini and automated content summarization and description generation.
• Text analytics: Built a three-class sentiment classifier and named entity recognition pipelines with Hugging Face Transformers.

BERT-Based Nepali Grammatical Error Detection and Correction Leveraging a New Corpus

Sumit Aryal and Anku Jaiswal (2024). Presented at IEEE INSPECT-2024, ABV-IIITM, Gwalior, India, December 07-08, 2024.

Created an 8.1-million-pair corpus across seven error types and fine-tuned MuRIL and NepBERTa for sentence-level grammatical error detection, achieving 91.15% and 81.73% accuracy, respectively, on 771,511 validation sentences with synthetically generated errors.

Nepali Grammar Correction

Sumit Aryal, et al. (2024). Undergraduate thesis, Pulchowk Engineering Campus, Institute of Engineering, Tribhuvan University.

Best Project Award, Pulchowk Campus · December 2024.

Research Agent

An agent that plans a research task, searches for information, reviews its findings, and writes an answer with citations. Built in Python with concurrent searches and configurable budgets and time limits.

Document Question Answering

A service for asking questions about uploaded documents and receiving answers with source citations. Uses OCR to extract text and combines keyword and semantic search, with FastAPI, PostgreSQL, and Qdrant.

Nepali Grammatical Error Correction

Nepali Grammatical Error Correction

As part of my bachelor thesis I built a BERT-based Nepali grammar correction (Nepali GEC) pipeline. I curated a large parallel corpus for the Nepali grammatical error correction task and implemented a system that ingests Nepali text, detects grammatical errors, and proposes context-aware corrections.

Travel Recommendation System

Built recommendation logic using custom TF-IDF and cosine similarity, plus item-based collaborative filtering. Filtered recommendations by travel preferences, budget, and trip duration, and assisted backend/frontend integration through Flask APIs.