Computer Engineering · IOE Purwanchal Campus · Nepal

$ init portfolio.py --mode=full

✓ Loading: Mistral-7B fine-tuned on Nepal Penal Code

✓ Loading: U-Net for satellite segmentation (mIoU: 0.674)

✓ Loading: 17 days of deep learning in public

✓ Status: Building AI for Nepal

YAMRAJ

ML & AI Builder

Fine-tuning LLMs, training U-Nets, writing research papers, and shipping code — all as a CS student from Nepal.

🇳🇵

See My Work GitHub Hugging Face

13.5GB

LLM Fine-tuned

4GB

GGUF Quantized

67%

U-Net Mean IoU

<24hrs

Community Adoption

Live Deployments

About Me

Not just studying
AI — shipping it

I'm a Computer Engineering undergraduate at Tribhuvan University, IOE Purwanchal Campus in Biratnagar, Nepal. While most students learn about ML from textbooks, I build production systems with it.

My flagship project — a fine-tuned Mistral-7B on Nepal's National Penal Code — was adopted by the open-source community within 24 hours of release, re-quantized into Q2–Q8 GGUF variants, and integrated into the llama.cpp ecosystem by community maintainers.

I presented my U-Net satellite segmentation paper at ICRTAI 2025, receiving direct feedback from Prof. Dr. Sudan Jha. I also built a multi-agent AI app, an Android legal assistant, and documented 17+ days of deep learning in public on LinkedIn.

Currently an AI intern at Planto AI, working on neural networks in production.

"Build things that matter for the place you're from."

— the philosophy behind every Nepal-specific AI project

Building in Public

17+

DL Days

509

Followers

Max Reactions

10+

Projects

Certifications

Conference

LocationBiratnagar, Kosi Zone 🇳🇵

CurrentAI Intern @ Planto AI

GitHub@yamrajkhadka

HuggingFace@yamraj047

LinkedInyamraj-khadka

Open ToML Internships · Research Collab

Projects

Things I've Built
& Shipped

Featured Project · 01

Nepal Legal AI System

End-to-end legal AI for Nepal's National Penal Code. Built the full pipeline: PDF extraction → hierarchical chunking → embeddings → RAG → Mistral-7B fine-tuning → GGUF quantization → 4 live deployments + Android app. The model was adopted by community maintainer mradermacher within 24 hours and re-distributed in Q2–Q8 formats.

LLM Fine-tuning Mistral-7B RAG + FAISS GGUF Quantization FastAPI React Native 4 Live Apps Community Adopted

🤗 FP16 Model → GGUF Model GitHub

Model Specifications

Base ModelMistral-7B

Original Size13.5 GB (FP16)

Quantized (Q4_K_M)4.07 GB

Community VariantsQ2–Q8, IQ, f16

Embedding Modelall-mpnet-base-v2

🚀 Community Impact

Within 24 hours, community maintainer mradermacher re-quantized and distributed the model into the full llama.cpp ecosystem. Runs on consumer CPU hardware — no GPU needed.

U-Net Land Cover Segmentation

Built U-Net from scratch in TensorFlow. Composite loss (Focal Tversky + weighted CCE). Deployed Streamlit app. Presented at ICRTAI 2025. Mean IoU: 0.674.

Published TensorFlow Live App

View Project

HerAI — Multi-Agent System

5 specialized LangGraph agents (Mood, Memory, Romantic, Surprise, Safety). RAG with FAISS for memory retrieval. Shows real multi-agent orchestration skills.

Agentic AI LangChain LangGraph

View Project

GGUF + RAG Legal Assistant

Real-time streaming via WebSocket. Bi-encoder + cross-encoder reranking. FAISS IVF index. Query LRU cache. Citation-format responses. Sub-second latency.

RAG WebSocket Live

View Project

RNN / GRU / LSTM / BRNN From Scratch

All four recurrent architectures implemented in pure PyTorch — no nn.RNN shortcuts. Built from the exact mathematical diagrams. Training loop + text generation.

PyTorch From Scratch Deep

View Code

MLOps: Trip Duration Prediction

MLflow experiment tracking, Optuna hyperparameter search, model registry, preprocessor serialization. Full reproducible pipeline on NYC taxi data.

MLOps MLflow Optuna

View Project

Nepal Legal Android App

Native Android app (Expo + React Native) powered by production GGUF API. Chat UI, chat history, auth. Includes document scanner with auto-crop. No on-device model.

React Native TypeScript APK Live

Download APK

Stack

What I Work With

I didn't just read about these tools — I used them to build real systems that run in production. Every item here maps to a project you can click on above.

17+

Days of Deep Learning in Public
Documented every concept from Perceptrons to U-Net on LinkedIn — the grind is real.

🧠 Deep Learning ▶

TensorFlow / Keras PyTorch CNN / U-Net RNN / LSTM / GRU Transfer Learning Albumentations Custom Loss Functions Regularization (L1/L2/Dropout)

🤖 LLMs & NLP ▶

LLM Fine-tuning Mistral-7B / LLaMA GGUF Quantization SentenceTransformers FAISS RAG Pipelines Instruction Tuning llama-cpp-python Groq API

🔗 Agents & Orchestration ▶

LangChain LangGraph Multi-agent Systems State Graph Routing Prompt Engineering

⚙️ MLOps & Deployment ▶

FastAPI Streamlit MLflow Hugging Face Spaces Optuna Docker WebSocket React Native / Expo

Learning in Public

The Grind Log

While studying CS, I documented 17+ consecutive days of deep learning on LinkedIn — building something real every single day, from scratch perceptrons to full CNN architectures. This isn't just coursework — it's obsession.

Perceptron

Built a Perceptron from scratch in Python. Visualized decision boundaries. Identified XOR limitations.

Shipped

MLP + Forward Prop

Implemented forward propagation manually in NumPy and with TensorFlow/Keras side-by-side.

Shipped

Backpropagation

Implemented backprop math step-by-step: δ[L]=A[L]−Y, chain rule, parameter updates.

Shipped

Optimizer Comparison

Ran SGD, Momentum, RMSprop, Adam across shallow/deep networks at 1K/10K/50K samples.

Shipped

CNN + 98.46% MNIST

Built CNN with conv, pooling, dense layers. Achieved 98.46% accuracy on handwritten digits.

Shipped

Transfer Learning VGG16

Fine-tuned VGG16 on CIFAR-10, frozen base layers. Achieved 73.13% with minimal compute.

Shipped

U-Net Architecture

Studied semantic segmentation → started applying U-Net to Nepal farmland identification from satellite.

Shipped

→

Then it escalated

From 17 days of learning to publishing a conference paper. From LeNet to fine-tuning Mistral-7B. The grind continues.

Ongoing

Research

Published Work

📄 Peer Reviewed · ICRTAI 2025 · Nepal

Land Cover Segmentation from Satellite Imagery Using U-Net with Custom Loss and Morphological Postprocessing

Yamraj Khadka · IOE Purwanchal Campus · Presented June 28–29, 2025

Proposes a U-Net architecture for 7-class pixel-wise land cover classification on the DeepGlobe dataset. Introduces a composite loss function (60% Focal Tversky + 40% weighted categorical cross-entropy) to address extreme class imbalance. Applies morphological and median filter postprocessing for cleaner segment boundaries. Trained with Albumentations augmentation pipeline. Deployed as a Streamlit web app with interactive visualization.

0.674

Mean IoU

Land Classes

ICRTAI
2025

Conference

Urban
Land

0.72

Agriculture
Land

0.72

Range
Land

0.35

Forest
Land

0.72

Water

0.66

Barren
Land

0.57

Unknown

0.98

Contact

Let's
Work
Together

I'm looking for ML internships and research collaborations. Especially excited about AI for low-resource, developing contexts. Let's build something that matters.

What I bring

▶

I ship things. Every project here is live, deployed, and usable — not just notebooks.

▶

Full-stack ML. From PDF parsing to LLM fine-tuning to Android app to production API.

▶

Research depth. Conference paper, custom loss functions, architectural choices with justification.

▶

Learning velocity. From perceptrons to Mistral fine-tuning in under a year of focused building.

▶

Nepal-first mission. I build AI that serves underrepresented languages and legal systems.

YAMRAJ

Not just studyingAI — shipping it

Things I've Built& Shipped

What I Work With

The Grind Log

Published Work

Models I've Released

Not just studying
AI — shipping it

Things I've Built
& Shipped