Kushal Dahal

Computer Science (AI) · Erasmus Mundus applicant

Kushal Dahal

Python backend developer from Nepal with a completed Bachelor of Information Technology. I build and evaluate LLM systems and generative models, and I am applying for a Master’s in AI/ML.

View projects Download CV
University
Tribhuvan University, Nepal
CGPA
3.44 / 4.0
Working as
Python Backend Developer, since Nov 2024
Focus
LLM agents, diffusion models

About

I completed a Bachelor of Information Technology at Tribhuvan University in 2025 and work as a Python backend developer at Volcussoft in Kathmandu, where I analyze Nepal Stock Exchange data for a financial analytics platform.

Alongside work I run independent research projects on code-generation reliability in LLMs, diffusion-based image generation and image captioning. I want to study AI/ML at master’s level, with a focus on LLM-based agents and generative models.

Research interests

  • LLMs and LLM-based agents
  • Automated bug detection and self-repair
  • Diffusion and score-based models
  • Vision-language and multimodal learning
  • Parameter-efficient fine-tuning (LoRA, adapters)

Target programmes

  • Programme name 1
  • Programme name 2

Projects

LLM Code-Generation Reliability: Sampling Strategies vs. Self-Repair

2026 · Independent research · Kaggle / GitHub

  • Built an evaluation pipeline that separates the effect of few-shot prompting, best-of-N sampling and two self-repair strategies on pass@1 accuracy.
  • Wrote a sandboxed, timeout-protected execution harness with an automated bug taxonomy: syntax, name, type, assertion and timeout errors.
  • Compared raw error-message feedback with diagnostic expected-vs-actual feedback. About 12% of failures were algorithmic reasoning errors that neither strategy fixed, which separates surface-level from logic-level bugs.
  • Qwen2.5-Coder
  • MBPP
  • pass@1
  • Self-repair
  • Python

Project link: add URL

Creative Image Generation with Diffusion Models

2026 · Paper reproduction · Kaggle

Sample output 1
Sample output 2
Loss curve or baseline comparison
Caption placeholder: say what the images show, for example creative outputs next to the unmodified Kandinsky 2.1 baseline.
  • Reproduced arXiv:2601.22125 (Song & Elgammal, Rutgers, 2026) on Kaggle T4/P100 with Kandinsky 2.1. The framework uses PCA-reduced Gaussian density estimation (768 to 50 dimensions), a differentiable negative log-likelihood creativity loss, and joint token and LoRA (rank 4) optimization on the diffusion prior’s text encoder.
  • Added dynamic anchor-loss switching and negative-cluster directionality control to limit out-of-domain drift and steer away from unwanted creative directions.
  • Scaled the paper’s 5000-sample A100 pipeline down to a 16 GB T4 with 500 prior samples, decoder offloading and a CLIP-threshold semantic checker in place of an MLLM validator. Novel outputs take under 2 minutes through a Gradio interface with checkpointing and loss visualization.
  • Kandinsky 2.1
  • LoRA
  • Diffusers
  • PyTorch
  • Gradio

Project link: add URL

Image Caption Generation with Attention Mechanisms

2024–2025 · Kaggle

  • Started from my final-year project on Flickr30k (30k image-caption pairs), a DenseNet201 encoder with a BiLSTM decoder. I then extended it on my own with a multi-head attention (Transformer-style) decoder.
  • Evaluated with BLEU-1 to 4, METEOR, ROUGE, CIDEr and SPICE, and checked generalization on out-of-distribution images.
  • DenseNet201
  • BiLSTM
  • Multi-head attention
  • Flickr30k

Project link: add URL

Agentic Chatbot with Retrieval-Augmented Generation

2024 · GitHub

  • Built a knowledge-grounded question-answering system that combines an LLM with document retrieval. LangChain handles orchestration, and FAISS with sentence transformers powers semantic search over a PDF processing pipeline.
  • Added task automation, including document querying and natural-language instruction following.
  • LangChain
  • FAISS
  • Sentence transformers
  • RAG

Project link: add URL

Education & Experience

Experience

Nov 2024 – Present

Python Backend Developer

Volcussoft, Kathmandu · Laganisutra, a financial data analytics platform

  • Analyzed large-scale NEPSE (Nepal Stock Exchange) floorsheet data with Pandas to find unusual transaction patterns, including wash trading and broker pairing behavior.
  • Ran exploratory analysis on market datasets to surface trading trends and anomalous activity.
  • Built data pipelines from market APIs and cut ClickHouse query time by 75% with vectorized operations.
  • Implemented pattern-detection algorithms for Head & Shoulders, Double and Triple Tops, Channels and Triangles. Prototyped an LLM-assisted pipeline for automated financial summaries, served through Django REST APIs.
  • Cut key API latency from 300 ms to 60 ms with Redis caching, and built asynchronous WebSocket services for live trading alerts and indicator updates.
  • Deployed a YOLO11m model as a FastAPI inference service for a live insurance survey application.
  • Mentored and onboarded a backend engineering intern.
  • Currently building a RAG system for multiple companies using Gemini and a FAISS vector store.

Education

2020 – 2025

Bachelor of Information Technology

Tribhuvan University, Nepal · CGPA 3.44 / 4.0

Final-year project: Image Caption Generator (DenseNet201 encoder with BiLSTM decoder, Flickr30k).

Enrolled 2020. University delays in exams and results pushed the final result to 25 September 2025. The Dean’s Office confirms this in a letter dated 27 January 2026, available on request. Convocation: December 2026.

2019

Higher Secondary Education (+2, Science)

Kathmandu Model College, Nepal

Skills

Languages and libraries
Python, SQL, NumPy, Pandas, SciPy, Scikit-learn
Deep learning
PyTorch, TensorFlow / Keras
ML and AI
Transformers (Hugging Face), Diffusers, FAISS, LangChain, Accelerate, PEFT
Backend and data
Django, Django REST Framework, FastAPI, Flask, WebSockets, PostgreSQL, MySQL, MongoDB, Redis, ClickHouse, ETL pipelines
Tools and platforms
Git, Docker, Linux, Jupyter Notebooks, Celery, RabbitMQ, Kaggle
Core competencies
LLM evaluation and prompting, code-generation benchmarking, computer vision, diffusion models, Vision Transformers, attention mechanisms, transfer learning, RAG systems

Certificates and training

  • Data Analysis with Python, freeCodeCamp
  • Machine Learning with Python, freeCodeCamp
  • Intro to Deep Learning, Kaggle Learn

Languages

  • Nepali: native
  • English: IELTS Academic overall 7.5 (minimum band 6.5)

Contact

I am happy to talk about Erasmus Mundus programmes, research and collaboration. Email is the best way to reach me.

Ask about me

Answers are written by an AI that reads my CV. It can make mistakes, so check anything important with my CV or by email.

Email Kushal.775424@bkmc.tu.edu.np
Email (personal) kushal.dahal.234@gmail.com
LinkedIn linkedin.com/in/kushal-dahal
GitHub github.com/kushaldahal0
Location Changunarayan, Bhaktapur, Nepal