About

I want to develop machine learning systems that resist attacks, stay reliable when data shifts, and help defend the software they run in.

I have a B.E. in Computer Engineering from the Institute of Engineering, Tribhuvan University. For my final year research project, Re‑Cyclo, I trained YOLOv8 to recognise household waste and connected it to a local LLaMA 2 model that suggests ways to reuse each item. Most of what I learned came from the experiments. Cleaning the data helped more than training for longer.

I joined BlueStone Africa in 2024 as its first developer. Today I lead the team behind its asset management platform, a set of Go services on Azure with a mobile app that technicians use in the field. Working on authentication and access control there is what pulled me toward security research.

Timeline

  1. May 2025

    Promoted to Senior Software Engineer at BlueStone Africa.

  2. Jul 2024

    Completed my B.E. in Computer Engineering at the Institute of Engineering, Tribhuvan University.

  3. Jun 2024

    Joined BlueStone Africa as its first developer.

  4. Mar 2024

    Finished Re‑Cyclo, my undergraduate research project on waste detection and reuse recommendation.

Research interests

Two questions interest me most. How do deployed models get attacked, and can machine learning help catch attacks on the kind of systems I already build?

  1. Adversarial Machine Learning

    Adversarial inputs, poisoned training data and stolen models. I want to understand these attacks well enough to build defences that still work after a model is deployed.

  2. Out-of-Distribution Robustness

    Models that keep working when the data changes without any attacker involved, such as noisy inputs or conditions never seen in training, and honest ways to measure that.

  3. ML for Intrusion Detection

    Learned detectors for intrusions and unusual behaviour in networks and applications, and how to make them harder for an attacker to slip past.

  4. Security of ML Systems

    Many attacks on a deployed model go through the API, the data pipeline or weak access control around it. That surrounding system is what I work on every day.

Research experience

Re-Cyclo

May 2023 – Mar 2024

Object detection and LLM-based reuse recommendation for household waste

Dept. of Electronics and Computer Engineering, Pashchimanchal Campus, Tribhuvan University

Most waste apps tell people how to recycle or throw something away. Re‑Cyclo looks at a photo of an item and suggests how to reuse it instead, which sits higher in the waste hierarchy.

Phone photo

Taken by a household user

YOLOv8 detection

19 waste classes

Local LLaMA 2 (7B)

Runs offline via Ollama

Reuse suggestions

Plus clustered pickups

Figure 1. The Re‑Cyclo pipeline, from a photo to reuse suggestions, with the language model running locally.
Waste classes
19
Controlled experiments
11
Datasets combined
4
LLM run locally
7B
  • Built the YOLOv8 detection pipeline for 19 household waste classes, combining four annotated datasets and adding augmentation for messy real-world photos.
  • Ran 11 controlled experiments across datasets, model sizes, training lengths and transfer-learning setups, which showed that careful data curation improved results more than longer training.
  • Connected detections to a local LLaMA 2 (7B, via Ollama) so reuse suggestions work fully offline, and designed centroid-fixed clustering to assign pickups to collectors.
  • Shipped the full system as open source.
YOLOv8LLaMA 2OllamaPythonFlutterNode.jsReact
Source code

Publications

  1. ManuscriptPublication 1

    Re-Cyclo: An AI-Powered Waste Management System Using YOLOv8 and Large Language Models

    Saroj Poudel, Nabin Lamichhane, Nishan Paudel, Pujan Lamichhane, Tilak Khatri

    Manuscript in preparation, 2025

    An end-to-end system that detects household waste with YOLOv8, generates reuse suggestions with an offline LLaMA 2 model, and clusters pickup requests to assign waste collectors.

    Code

Education

Bachelor of Engineering in Computer Engineering

Nov 2019 – Jul 2024

Institute of Engineering, Pashchimanchal Campus, Tribhuvan University

Pokhara, Nepal

Relevant coursework. Artificial Intelligence, Distributed Systems, Operating Systems, Computer Networks, Database Management Systems, Computer Organization & Architecture, Enterprise Computing.

Skills

Languages
Python, Go, TypeScript, C++, Dart, SQL
ML & Data
PyTorch, YOLOv8 (Ultralytics), scikit-learn, OpenCV, NumPy, Pandas, LLMs (LLaMA 2, Ollama)
Systems
Microservices, gRPC, REST, PostgreSQL, Redis, Docker, Kubernetes, Azure, CI/CD, Linux
Research
Experimental design, Ablation studies, Performance evaluation, Literature review, LaTeX

Industry experience

Most of my research questions started at work. Fixing authentication bugs and watching a multi-tenant platform grow showed me where real systems break.

Senior Software Engineer, BlueStone Africa

May 2025 – Present · Remote

Cloud-based enterprise asset management platform for facilities and property portfolios

  • Lead a team of developers building a multi-tenant platform for two enterprise clients (about 1,000 users).
  • Designed Go microservices linked by a message broker, keeping each client's configuration and data separate.
  • Led the design of PostgreSQL partitioning and Redis caching so queries stay fast as client data grows.
  • Run services on Kubernetes, with CI/CD pipelines that test changes before they deploy to Azure.
  • Set up monitoring and alerting, introduced role-based access control and fixed authentication vulnerabilities.
  • Built the mobile app for asset verification and maintenance, with offline mode, QR scanning and planned inspections.
  • Review code, mentor developers and plan upcoming work with product and leadership.
GoMicroservicesPostgreSQLRedisDockerKubernetesAzureCI/CDFlutter

Software Engineer, BlueStone Africa

Jun 2024 – Apr 2025 · Remote

  • Built the platform from scratch as its first developer, using React, Node.js and MongoDB.
  • Shipped features in short release cycles driven by feedback from early facilities teams.
  • Launched the platform in production for its first enterprise clients.
MongoDBExpressReactNode.js

Academic projects

Larko, Distributed Queue Management System

Feb 2023 – May 2023

  • Designed FIFO scheduling with capacity planning based on Little's Law, real-time WebSocket updates and QR code check-in, handling 200 concurrent sessions in load testing.
FlutterDjangoPostgreSQLReactWebSockets

Contact

Faculty and research groups working on AI security or machine learning are welcome to get in touch. Email is the fastest way to reach me.

Email sarojpoudel201@gmail.com, or send a message below.

0/1000