Samir Poudel · MTSU
Software Engineer & AI Researcher
PhD Student in Computational & Data Science · Middle Tennessee State University
I work in both industry and academia. In industry, I build APIs, data pipelines, and AI tools that people use every day. In research, I study how to make machine learning private, secure, and robust, with a focus on federated learning. Right now I’m a Graduate Research Assistant at MTSU.
Focus Applied AI · Data engineering · Federated learning · ML security
I’m from Pokhara, Nepal, and based in Murfreesboro, Tennessee. I spent more than eight years building software and data systems — at Urja Tech, TwoAce Solutions, and DataInsight — before and during graduate school.
I’m now a PhD student in Computational and Data Science at Middle Tennessee State University (expected May 2028). I completed my MS in Computer Science at MTSU in December 2025 and my BS in Software Engineering in 2022. My research turns into working code, and my engineering work keeps my research practical.
Software Engineering Intern · Backend & Data
I built HomeInsight Core, a service that serves 2.4 million Tennessee property records.
Full Stack Developer
I built web apps from the database all the way to the screen.
Software Engineer · Data & AI
I turned slow, manual reporting into automatic reports using Python and SQL.
Graduate Research Assistant
I research how to make machine learning private, secure, and fast. So far that has led to 17 published papers, 3 where I am the first author.
I have published 17 papers on private and secure AI, AI for health, and big data systems. I am the first author on 3 of them.
IEEE AIIoT 2025
Six regions trained one model to spot forest loss from satellite photos without sharing their data. The better method reached 0.85 accuracy (IoU), up from 0.77, and treated every region more fairly.
IEEE TPS 2025
A defense that catches hidden "backdoor" attacks when many computers train one model together, keeping attack success below 15%.
Journal of Marketing Analytics 2023
Using machine learning to predict cryptocurrency prices and how much they swing.
Training AI on data that never leaves its owner, and stopping attackers who try to poison the model.
Using AI that reads code in several forms at once to find security holes in blockchain contracts.
Predicting hospital stay length, detecting stress from wearables, and reading brain scans.
Measuring and speeding up the tools that move and process millions of rows of data.
01Applied AIPersonal
A trust model that cut the effect of fake-review campaigns on product ratings by 93%.
Built on a Databricks lakehouse with MLflow tracking each training run. I tested it on real reviews with fake "review farm" attacks mixed in; the 93% is how much less the ratings moved after the model stepped in.
02Applied AIPersonal
An AI assistant that can search my email, find documents, set reminders, and help with finances.
It decides which tool to use, stops itself from looping forever, skips repeated calls, and streams its answer as it types. Works on the web and in Telegram.
03Data engineeringIndustry
A service that serves 2.4 million Tennessee property records, built at DataInsight.
It powers tenant screening with credit, background, and job checks, plus payments, repair requests, and messages for more than 100 properties.
04Data engineeringResearch
Pipelines for 10 million+ rows, and a published test of which big data tool runs faster and uses less memory.
Compared Apache Spark with Apache Airflow (IEEE CCWC 2025), and sped up data prep and model training by running them in parallel with MPI4Py (Machine Learning with Applications, 2023).
05Data engineeringPersonal
Tracks U.S. immigration (I-140) case updates automatically and shows the trends on a dashboard.
It checks case numbers on its own, saves the results in PostgreSQL, and shows them through a FastAPI backend and a React dashboard.
06ResearchResearch
Catches "backdoor" attacks when many computers train one AI model together (first author, IEEE TPS 2025).
It looks at what each computer’s model has learned and flags the poisoned ones, keeping attack success below 15% while the model stays accurate.
07ResearchResearch
Finds forest loss in satellite photos while each region keeps its own data (first author, IEEE AIIoT 2025).
Six regions plus one noisy "trouble" source trained a shared image model on 2015–2023 Sentinel-2 photos. The better method (Krum) reached 0.85 IoU vs. 0.77 and was fairer across regions.
I’m currently a Graduate Research Assistant at MTSU. I’m always happy to talk about research, collaborations, and summer internships.
spoudel04@gmail.com