Pramish Thapa
I am
About
AI researcher and software engineer building trustworthy, reliable AI systems.
AI Researcher & Software Engineer
First-author manuscript in preparation for MDPI Sensors on cross-dataset benchmarking of ML degradation models, backed by nearly three years of production engineering experience in Java and Spring Boot on Google Cloud Platform for Equifax (via Oncore Corporation).
- Role: AI Researcher & Software Engineer
- Research: Trustworthy AI, LLM agent security, uncertainty quantification
- Engineering: Java, Spring Boot, GCP
- Location: United Kingdom
- Degree: MSc Software Engineering (Distinction)
- University: University of West London
- Availability: Open to opportunities
My research focuses on trustworthy and reliable AI, the security of LLM agents, reproducible ML evaluation and uncertainty quantification. I am currently studying prompt-injection robustness of LLM agents. Before research, I delivered credit risk and identity verification systems for Equifax, and I hold an MSc in Software Engineering with Distinction from the University of West London.
Tech Stack
Languages, frameworks, libraries and platforms I have studied and deployed.
Languages
Front End
Back End
Databases & Messaging
Cloud & DevOps
Machine Learning
Tools
Curriculum Vitae
A summary of my experience and education.
Pramish Thapa
AI Researcher & Software Engineer · Trustworthy AI · LLM Agents · Java · GCP
AI researcher and software engineer. First-author manuscript in preparation for MDPI Sensors on cross-dataset benchmarking of ML degradation models, with emphasis on statistical robustness and uncertainty quantification. Production microservices experience in Java and Spring Boot on Google Cloud Platform for Equifax. MSc Software Engineering with Distinction from the University of West London. Research interests: trustworthy and reliable AI, security of LLM agents, reproducible ML evaluation, and uncertainty quantification.
Professional Experience
Independent Research and Upskilling
Self-directed · London, UK Oct 2024 - PresentJava Developer
Oncore Corporation · client: Equifax Sep 2021 - Jun 2023- Built and maintained Java / Spring Boot microservices on Google Cloud for Equifax credit-risk and identity-verification systems.
- Designed asynchronous data pipelines (Cloud Pub/Sub, Firestore) deployed through Jenkins CI/CD, with 95%+ test coverage on owned services.
- Built services for a regulated financial-services client; presented technical work to client stakeholders in 50+ review meetings.
Java Intern
InfoxIT Pvt. Ltd Jan 2020 - May 2021- Developed backend ERP modules in Java, Spring and Hibernate for banking clients.
- Designed MySQL schemas and optimised queries based on user feedback.
Education
MSc Software Engineering
Distinction Sep 2023 - Sep 2024Dissertation (85/100): digital twin and IoT pipeline for building energy monitoring; Random Forest model achieving R² of 0.999 on held-out data.
Bachelor of Computer Information Systems (BCIS)
CGPA 3.31/4.0 Sep 2016 - Jul 2021Modules graded A include Major Project, Data Structures and Algorithms, and Mobile Computing.
Further Training
- Neural Networks and Deep Learning - DeepLearning.AI (2026)
- Full Stack Developer Traineeship - IT Career Switch (2025 - 2026)
- Full-Stack Engineer Career Path - Codecademy (2025)
Languages
English (fluent) · Nepali (native) · Hindi (fluent) · Spanish (basic)
Research and Publications
No Single Model Dominates: A Regime-Aware, Cross-Dataset Benchmark of Degradation Models for Lithium-Ion Battery Remaining Useful Life
First author In preparation- Seven models including CNN-LSTM, Gaussian process regression and XGBoost benchmarked across the NASA PCoE and Oxford datasets under a chronological evaluation protocol.
- Bootstrap confidence intervals, Wilcoxon tests with Holm correction, rolling-origin validation and leave-one-cell-out jackknife.
- Interactive research dashboard (React / Recharts) reproducing every figure and metric in the manuscript: battery-rul-dashboard.vercel.app
Prompt-Injection Robustness of LLM Agents
Ongoing- Reproducing the AgentDojo benchmark on current Claude models, measuring task utility and attack success: github.com/Pramish17/agent-reliability
CNN-LSTM Remaining Useful Life Prediction (NASA C-MAPSS)
- CNN-LSTM remaining-useful-life prediction on NASA C-MAPSS turbofan degradation data, RMSE 14.84, with SHAP interpretability analysis.
Technical Skills
Python, PyTorch, TensorFlow, scikit-learn, pandas, NumPy, SHAP; CNN, LSTM and hybrid architectures; Gaussian process regression, XGBoost, Random Forest
Java, Spring Boot, REST APIs, Google Cloud Platform, Docker, Jenkins CI/CD, Git, SQL (MySQL, PostgreSQL), JavaScript / React
Portfolio
Selected projects. Click a card for details - descriptions open in a modal, not a new page.
Credit Risk Scoring Service
Company Directory
MeroJobRadar
Smart Building Energy Prediction
Battery RUL Dashboard
Gazetteer
Earth Letters
Contact
Have a role or a project in mind? Send me a message.