Kasmik Regmi

About

Final-year Computer Engineering student specializing in AI systems, deep learning research, and high-performance backend engineering. Experienced in client-facing LLM research and building AI/backend pipelines.

Experience

• ML & Backend Engineer
  Himalogic
  2026 – 2026
• AI Fellow
  Konnect Craft
  2026 – 2026
• Freelance LLM Researcher
  AI Plans
  2024 – 2025

Education

• B.E. Computer Engineering
  Kathmandu University
  Computer Engineering · 2022 – 2026

Skills

• SQL
• C
• FastAPI
• Python
• JavaScript
• Django
• NumPy
• Pandas
• scikit-learn
• OpenCV
• React
• PyTorch
• TensorFlow
• LangChain
• API Integration Understanding
• Unity
• LangGraph
• MCP
• Vue

Projects

• Hybrid RAG & Knowledge-Graph Engine (Fellowship Project) ↗
  LangChain, FastAPI, Neo4j, LLM, Ollama

  Engineered the retrieval and ingestion core of a document-QA product with PDF parsing, vision-based entity extraction into a Neo4j knowledge graph, dense vector search, graph traversal, hierarchical clustering, reciprocal rank fusion, and cross-encoder reranking.

• AI DevOps Agent ↗
  LangGraph, MCP, FastAPI, PostgreSQL, Graphify

  Built an agent that can view a codebase and execute natural-language GitHub operations from Telegram, with durable human-in-the-loop approval using PostgreSQL checkpointer support.

• Dense-to-MoE Upcycling ↗
  PyTorch, Hugging Face, Mixture-of-Experts, LLM

  Upcycled a pretrained dense LLM into a sparse Mixture-of-Experts model, reducing validation perplexity and improving training stability.

• Cairn: Voice-Driven Mental Health Support Platform ↗
  FastAPI, Supabase, wav2vec2, LLM, NLP

  Built a platform that turns daily voice memos into private emotional profiles and forms balanced support circles matched on life context. Used wav2vec2 for acoustic embeddings, and LLM to track the linguistic features.

• Satellite-Based Flood Prediction with a Late-Fusion Model ↗
  PyTorch, Google Earth Engine, Pandas, LSTM, MLP

  Curated a georeferenced dataset of 10,088 observations, with preprocessing designed to break spatial autocorrelation. Architected and trained a three-branch late-fusion network: BiLSTM + MLP + LSTM, with stratified sampling and GroupShuffleSplit enforcing spatial train/test separation. Achieved AUC-ROC 0.853 and 97.7% flood recall on previously unseen locations.

• SFaceX: Adaptive Face Lock System ↗
  Python, OpenCV, Computer Vision, Scikit-learn

  Built a lightweight facial recognition system with adaptive learning, preprocessing, and behavior-learning for unlock preferences.

Courses & certifications

• AI Fundamentals · DataCamp · 2026

🏆 Achievements & awards

• Track Winner (AR/VR & Immersive Reality) · 2026

  KU Hackfest winner in the AR/VR & Immersive Reality track.

• JunctionX Kathmandu Hackathon Finalist · 2026

  Finalist in the JunctionX Kathmandu Hackathon.

🎯 Hobbies & interests

• R&D
• AI
• Backend Engineering
• Agentic Systems