Skip to content
Available for ML engineering roles & research

I turn machine learning into dependable software.

I'm Khagendra Luitel, a Machine Learning Engineer based in Sunsari, Nepal. I work across the full lifecycle: training PyTorch models, structuring data pipelines, and serving low-latency APIs.

Curriculum Vitae ↗GitHub ↗
Deep Learning

Training loops, architectures and evaluation, in PyTorch. The base the rest of the work stands on.

PyTorchTraining loops
LLMs

Transformers, tokenisation and retrieval, studied on top of the modelling base rather than beside it.

TransformersRAG
Computer Vision

Image models and the visual side of a pipeline. Stated as a direction, because the CV records no work here yet.

Image modelsVision
Agentic AI

Tool use, planning loops and orchestration around a model, which is where the serving work is headed.

Tool usePlanning
NPT (UTC+5:45)
Core Toolkit
  • Python
  • PyTorch
  • Scikit-learn
  • XGBoost
  • NumPy
  • Pandas
  • FastAPI
  • PostgreSQL
  • SQL
  • n8n
  • Git
  • Linux
Featured Work

Production machine learning,engineered end to end.

Selected engineering projects showcasing model training, dataset engineering, and live API deployment.

Index03 systems · most recent first
REQUESTPOST /predictFASTAPIScikit-learn estimator{ }
FastAPIScikit-learnPython
ML System · Model Serving2026

Insurance Premium API

Trained a model on insurance data, then wrapped it in a FastAPI service so it can actually answer requests instead of sitting in a notebook.

Problem Statement

The model worked fine locally. But models in notebooks don't help anyone - I needed it running somewhere it could take real requests and return real predictions.

Key Implementation
  • Built the API with FastAPI - loads the model on startup, takes a policy record, returns a premium prediction.
  • Tested it with curl and Postman before calling it done. The model isn't fancy (Random Forest), but the pipeline works.
Architecture Pipeline
  1. Policy dataset
  2. Scikit-learn
  3. Joblib dump
  4. FastAPI
  5. JSON endpoint
Technical Specifications
Endpoint
POST /predict
Input
JSON policy record
Output
Premium estimate (float)
Latency
~50ms per request
Detailed Case Study
Code is private for now. Happy to share on request.
DISTRIBUTIONTREND
PythonPandasMatplotlib
Data Analysis · Visualisation2026

HR Attrition Analysis

Raw HR data that nobody was reading. I cleaned it, found where people were actually leaving, and gave the team charts instead of SQL queries.

Problem Statement

HR had numbers but no story. The attrition was there in the data, but getting a simple answer meant wrestling with SQL joins. I cleaned the data and made it readable.

Key Implementation
  • Pulled the raw export, fixed typos and missing values, grouped by department and tenure.
  • Found that attrition clusters in the first 6 months and specific departments - things the raw numbers hid.
Architecture Pipeline
  1. Raw HR records
  2. Pandas cleaning
  3. Exploratory analysis
  4. Matplotlib charts
  5. CSV summary
Technical Specifications
Data source
HR export (CSV)
Analysis
Attrition by dept, tenure, role
Deliverable
Charts + one-page summary
Finding
6-month cliff in attrition
Detailed Case Study
Notebook is on my machine. I'll walk you through it.
LINEAR REGDECISION TREERANDOM FORESTEVALUATIONR2MAERMSE
PythonSQLPandasScikit-learnSeaborn
ML Model Comparison2026

Freight Cost Prediction

Pulled invoice data, engineered one feature that mattered, then ran three models head-to-head to find what actually predicts freight costs.

Problem Statement

Nobody could answer 'what should this shipment cost?' The data was there but nobody had built anything to compare against - just raw invoices sitting in SQL.

Key Implementation
  • Started with correlation analysis - quantity, freight cost, invoice value. Freight per unit showed the strongest signal.
  • Ran Linear Regression, Decision Tree, and Random Forest side-by-side. Random Forest won on R².
Architecture Pipeline
  1. Invoice SQL dump
  2. Pandas cleanup
  3. Feature: freight per unit
  4. Compare 3 models
  5. Pick best (Random Forest)
Technical Specifications
Best model
Random Forest
R² score
0.87
Key feature
Freight per unit
Output
Predicted cost per shipment
Detailed Case Study
Can share the notebook. Just ask.

Two projects are private client/coursework implementations. Walkthroughs and code samples are available on request.

03 Case Studies
Profile OverviewSunsari, Nepal · NPT (UTC+5:45)
Focus
Machine Learning Engineer
PyTorch · Scikit-learn · FastAPI
Education
Bachelor of Computer Science
Madhan Bhandari Memorial Academy Nepal · Expected graduation 2026
Experience
Data Science Intern
Cloud Nepal Web · May to July 2026
Status
Looking for work
ML roles with real data
Career & Education

Experience, leadership,and academic foundation.

Internship experience, organizational leadership, and formal Computer Science coursework.

  1. EducationExpected 2026

    Bachelor of Computer Science

    Madhan Bhandari Memorial Academy Nepal, Urlabari

    • Expected graduation 2026
    • Coursework: Machine Learning, Artificial Intelligence, Probability & Statistics, Data Analysis & Visualization
  2. ExperienceMay to July 2026 · 3 months

    Data Science Intern

    Cloud Nepal Web

    • 3-month internship doing real data work - cleaning datasets, building models, and actually using them on company data.
    • Learned to move from 'it works in my notebook' to 'it runs on their servers'.
    • Python
    • Pandas
    • Machine Learning
  3. Leadership2026 to present

    Operations Lead

    CFC Koshi

    • Run operations for a student org. Get things done with limited resources - classic startup conditions.
    • Coordinate events, manage the team, make the trains run on time.

I've kept all of this as it appears on my CV.

03 entries
Open Source

Code and contributions,published on GitHub.

Live repositories synchronized with GitHub, updated hourly to reflect recent work and open source activity.

Repositories4 of 21 repositories
Public activity · last 13 weeks1 active day
Engineering Philosophy

How I approachmachine learning engineering.

My focus is on models that solve concrete problems, evaluate honestly, and run reliably in production environments.

I build machine learning systems designed for production. I work across the full pipeline: engineering dataset features, training and benchmarking model architectures, and deploying reliable inference services with PyTorch and FastAPI.

Most machine learning value is unlocked through disciplined fundamentals: cleaning and validating datasets, engineering features with domain context, establishing reliable baselines, and deploying via well-tested APIs. My background in cybersecurity also instills a strong security and robustness mindset: understanding failure modes makes systems significantly more resilient.

Training models

Getting real signal from messy data

Making them work

PyTorch, debugging loss curves, knowing when you're done

Shipping them

APIs, endpoints, making sure they don't break in production

The new stuff

Transformers, LLMs - learning what's next

Areas of Interest
  • Building things with ML
  • Model deployment
  • Deep learning
  • Data systems
  • Knowing how it breaks
Education

Bachelor of Computer Science

Madhan Bhandari Memorial Academy Nepal, Urlabari

Expected graduation 2026

Relevant Coursework
Machine LearningArtificial IntelligenceProbability & StatisticsData Analysis & VisualizationData Structures & AlgorithmsResearch Methodology
Certifications
  • Introduction to CybersecurityCisco Networking Academy
  • Foundation Level Threat Intelligence AnalystarcX
Technical Stack

Core competencies,tooling, and frameworks.

An interactive capability graph organized by engineering domain, backed by verified project evidence.

Drag to turn it. Select a node to see where that skill is used.

Node size is how many entries record that skill: a case study, the internship, a course, or something being studied right now. Every one of those entries is listed the moment you select the node.

GenAI, Agentic AI and RAG are marked as learning, because that's where they are today.

AI Architecture

Generative and agenticsystem design.

Investigating language models as reasoning cores, augmented with vector retrieval, external tools, and autonomous execution loops.

Intelligence Pipeline6 Core Stages
  1. 01
    InputPrompt · documents · signals
  2. 02
    LLMTransformer inference core
  3. 03
    ReasoningPlanning & decomposition
  4. 04
    ToolsRetrieval · vectors · APIs
  5. 05
    AgentAct · observe · iterate
  6. 06
    OutputGrounded, verifiable result

Generative AI

Model behaviour, prompting, evaluation

Agentic AI

Tool use, planning, autonomy loops

Large Language Models

Architecture & fine-tuning

Transformers

Attention, tokenisation, context

RAG

Retrieval, embeddings, grounding

AI Systems

Serving, orchestration, reliability

Actively studying these

Research & Literature

Foundational papersand ongoing study.

Key machine learning literature and deep-dive technical material I analyze to inform architectural design decisions.

Selected Literature2 Publications

  • Attention Is All You Need

    Vaswani et al., 2017

    Transformers
    Studying
    • Self-attention
    • Multi-head attention
    • Positional encoding

  • Build a Large Language Model (From Scratch)

    Sebastian Raschka

    LLM engineering
    Studying
    • Tokenisation
    • Embeddings
    • Attention
    • Transformer blocks
    • Pretraining
Get in Touch

Let's build something exceptional.

Available for ML engineering roles & research. If you have an engineering problem, an ML project, or a research direction to explore, I'd love to connect.

Send EmailCurriculum Vitae
GitHub

@hacker-194

LinkedIn

in/khagendra-luitel

Email

anjalluitel3@gmail.com

Sunsari, Nepal+977-9827060306