now:
studying: "BSc (Hons) Computing with AI, Islington (2028)"
working: "Junior AI Developer intern @ Aviyaan Tech (Mar 2026 — present)"
building: "ensemble models for traffic counts, road-damage segmentation + classification"
open_to: ["junior AI/ML roles", "freelance RAG + agent work"]
based_in: "Kathmandu, Nepal (UTC+5:45)"ML/AI engineer in Kathmandu. I build retrieval systems, agent loops and vision models, then write down what broke and why.
BSc (Hons) Computing with AI student at Islington College (graduating 2028), currently interning as a Junior AI Developer at Aviyaan Tech. I like work I can point at and say: that runs because of me.
Junior AI Developer (Intern) — Aviyaan Tech · Mar 2026 — present Building and experimenting with ML models — ensemble methods for traffic-count estimation, plus road-damage segmentation and classification — from data prep through training and evaluation.
Independent Builder — Personal & demo projects · 2024 — present Learning in public: RAG loops, agent harnesses, vision models, and a full-stack demo storefront (EcoVerda). No client work yet — available for freelance. Most of the work below was built here.
BSc (Hons) Computing with AI — Islington College, Kathmandu · 2025 — 2028 Undergraduate degree focused on AI and ML. Coursework and self-directed builds in PyTorch, RAG pipelines, agent tool-calling, and full-stack deployment.
+2 Science (Computer Science) — Global School of Science · 2024 — 2025 Higher secondary education in the Science stream with Computer Science.
| Project | What it does | Stack | Type |
|---|---|---|---|
| ai-research-agent — Ask questions over your PDFs, Markdown or text files. The agent routes between document search, web search, or both, and streams every tool call to the browser live instead of hiding the loop behind a framework. The whole loop is ~120 lines of plain Python. | Python FastAPI ChromaDB SSE |
 |
|
| small-agent-reliability — Nine open-weight models (1B–9B) scored as tool-using agents across 31 capability and 14 reliability tasks. Accuracy, consistency, robustness, failure recovery and refusal behaviour scored separately, not collapsed into one leaderboard. Reproduces on a laptop via Ollama. | Python Ollama pandas LaTeX |
 |
|
| pdf-chat-rag — Document Q&A with real CRUD: deleting a document removes its vectors, file and database row in one operation — the part most demos skip. Answers stream token by token over SSE. | Python FastAPI ChromaDB PyMuPDF |
|
|
| vehicle-image-classifier — ResNet-18 transfer learning on 400 images across four classes, labeled with my own annotation tool. Reports a confusion matrix and per-class accuracy next to the headline number, because one aggregate hides a model that's good at one class and guessing at another. Served behind FastAPI, containerised. | PyTorch ResNet-18 FastAPI Docker |
|
|
| EcoVerda — Personal demo storefront for sustainable products: cart persistence, credentials auth, orders and reviews on Next.js 16 + Prisma/SQLite. The Stripe client exists but isn't wired into checkout yet — tracked as a known issue in the repo. Live demo | Next.js TypeScript Prisma Stripe |
|
|
| ParkX — Smart parking across three layers: ESP32 sensor firmware, a FastAPI backend, and a dashboard polling bay state every 2 seconds. Bookings, anomaly alerts, and a local-Ollama-first, Groq-fallback chatbot. | ESP32 FastAPI Ollama |
|
More repos — experiments, coursework, and old tools (10)
- demand-predictor-ml — parking demand forecasting experiments with XGBoost and scikit-learn.
- health-guard-ml — early-stage health-risk classifier; still a stub with honest TODOs in the README.
- nnunet-road-cracks — crack-segmentation packaging over nnU-Net: one-click training and prediction scripts aimed at non-technical users.
- career-stability-analyzer — data exploration on job-stability survey data.
- ai-doc-summarizer — abstractive summarization API experiment (archived).
- qgis-batch-extraction — QGIS batch-processing scripts for geospatial layers (archived).
- vehicle-labeling-tool — my PyQt5 annotation app, used to label the vehicle classifier's 400-image training set (archived).
- todo-list-cli — my first Python CLI project (archived).
- js-calculator — weekend JavaScript calculator.
- MyProjects — old scratch repo of first experiments (archived).
Comfortable Python, FastAPI, RAG pipelines, ChromaDB, prompt and tool design, PyTorch, scikit-learn, pandas, Docker, Git and GitHub Actions, SQLite, REST APIs
Used on real projects Next.js, TypeScript, React, Tailwind, PostgreSQL, Prisma, Stripe, XGBoost, OpenCV, ESP32, REST polling, Linux and shell
Studying now CUDA and GPU profiling, GGUF quantisation, MLOps and experiment tracking, distributed training
| Stats | Streak |
|---|---|
Open to ML/AI engineering roles — remote or in Kathmandu.
