title: Andrej Janchevski - PhD Research
emoji: 📚
colorFrom: green
colorTo: gray
sdk: docker
app_port: 7860
pinned: false
license: mit
Andrej Janchevski — PhD Research
Live demos for the PhD thesis Scalable Methods for Knowledge Graph Reasoning and Generation (Andrej Janchevski, EPFL, 2025). The thesis is mirrored at infoscience.epfl.ch; a CV is rendered at /cv from docs/cvAndrejJanchevski.pdf.
The site is a single Django + Vue Docker container. The Django backend serves a stateless REST API at /api/v1/* and WhiteNoise serves the built Vue SPA on every other path — same origin, no CORS in production. Deployment target: Hugging Face Space Bani57/website → https://bani57-website.hf.space.
What's in the demos
| Demo | Thesis | What it does |
|---|---|---|
| COINs — KG reasoning | §3.1 | Link prediction and multi-hop / intersection query answering on FB15k-237, WN18RR and NELL-995, with six embedding algorithms (TransE, DistMult, ComplEx, RotatE, Q2B, KBGAT). |
| MultiProxAn — graph generation | §4.3 | Discrete denoising diffusion for molecular and synthetic graphs, with an outer Gibbs loop (MultiProx) that round-trips through /continue for live previews. |
| KG anomaly correction | §4.4 | Diffusion-based denoising of knowledge-graph subgraphs with a per-frame KG log-likelihood metric from the frozen COINs link ranker. |
Layout
.
├── Dockerfile, environment.yml, docker-compose.yml, entrypoint.sh, .dockerignore
├── README.md # this file (also the HF Space landing card)
├── src/backend/ # Django REST API
├── src/frontend/ # Vue 3 SPA
├── src/research/ # research code, imported via sys.path
├── scripts/ # one-off helpers (checkpoint upload, motif export)
├── docs/ # technical documentation (start at docs/README.md)
└── plans/ # implementation plans for major work
Quick start
Option A — Docker (production-equivalent)
cp .env.example .env
python -c "import secrets; print('DJANGO_SECRET_KEY=' + secrets.token_urlsafe(50))" >> .env
docker compose up --build
http://localhost:7860. First build is ~20–30 minutes (mamba env solve, GPU torch wheels, frontend bundle, ~5.4 GB checkpoint pull from HF Hub). Subsequent runs reuse layer cache and the checkpoints named volume.
Option B — local Python + Vite (dev)
mamba env create -n website_c -f environment.yml
mamba activate website_c
pip install --extra-index-url https://download.pytorch.org/whl/cu118 -r src/backend/requirements.txt
( cd src/backend && DJANGO_DEBUG=True python manage.py runserver 8000 ) &
( cd src/frontend && npm install && npm run dev )
Backend on :8000, Vite on :5173. The dev frontend env file points the SPA at the local backend.
Full instructions: docs/guides/local-development.md.
Documentation
The docs/ directory holds the full documentation set, organized by Diataxis document type:
docs/README.md— index and routing.docs/glossary.md— domain terms (COINs, MultiProxAn, query structures, SSE, inference lock, …).docs/explanation/— architecture, research methods, inference lifecycle.docs/reference/— API, SSE protocol, backend services, frontend modules.docs/guides/— local development, deployment.docs/api.yaml— OpenAPI 3.0.3 source of truth for the REST API.docs/postman/— Postman collection and environment for hand-testing.
The Django backend has its own README at src/backend/README.md and the Vue SPA at src/frontend/README.md. Both link back to docs/ for the deeper material.
Deploying
git push hf master:main to the Space remote. See docs/guides/deploy.md for the full procedure (Space secrets, checkpoint refresh on HF Hub, rollback, persistent storage option).
License
MIT for the website code. The PhD thesis and individual research methods retain their original publication terms; see the thesis PDF.