A document-conversion platform I've been building. Web UI + REST API + WeChat mini-program from a single Python backend. Started as a personal tool, grew into something a few teammates started using, so I tightened it up.

The reason for the WeChat/H5/Alipay version is that most users in our context aren't on desktop. Sharing a converted file via a tiny client is way more useful than asking them to log into a website.

• Auth: email + password, JWT (access/refresh), email verify, TOTP 2FA, password reset
• Files: upload, scan, store, stream back. S3/MinIO in prod, local disk in dev
• Conversion: ~20 source/target pairs across document, spreadsheet, image, data
• Backends used: LibreOffice, Pandoc, Pillow, WeasyPrint, Tesseract OCR
• Batch: zip up to 200 files, get one zip back
• Webhooks: HMAC-signed callbacks when a task finishes
• Rate limit, audit log, idempotency keys
• WebSocket for live progress
• i18n (zh/en) on the web side

cp .env.example .env
cp backend/.env.example backend/.env # if present; otherwise backend reads from .env at repo root
make keys # generates keys/jwt_{private,public}.pem
docker compose -f docker-compose.dev.yml up -d

API on :8000, web on :5173, MinIO console on :9001 (minio/minio123).

If you don't want Docker:

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
# start postgres + redis separately, or use docker compose just for those:
docker compose -f docker-compose.dev.yml up -d postgres redis minio
alembic -c backend/alembic.ini upgrade head
python -m backend.app.scripts.bootstrap_admin
uvicorn app.main:app --reload --app-dir backend

For the worker:

celery -A app.workers.celery_app worker -l info --app=app.workers.celery_app -C 2

backend/ FastAPI + Celery
frontend/ React + Vite + Tailwind + Zustand
miniprogram/ Taro (compiles to weapp / h5 / alipay)
deploy/ nginx, prometheus, grafana, k8s
docker-compose*.yml
Makefile

• POST /v1/auth/{register,login,refresh,logout}
• POST /v1/auth/password-reset/{request,confirm}
• POST /v1/auth/email-verify/{request,confirm}
• POST /v1/auth/totp/{setup,verify,disable}
• GET /v1/users/me, PATCH /v1/users/me
• POST /v1/files/uploads, GET /v1/files/{id}, GET /v1/files/{id}/download
• POST /v1/convert, GET /v1/tasks/{id}, POST /v1/tasks/{id}/{retry,cancel}
• POST /v1/batches, GET /v1/batches/{id}
• GET /v1/formats (full graph, frontend uses this to build the format picker)
• POST /v1/webhooks, etc.
• WS /v1/ws/tasks/{id} for progress

OpenAPI at /docs in dev.

A list so I don't forget, in priority order:

• backend/app/converters/image/image_converter.py — EXIF orientation handling for CMYK TIFF input is patchy. PIL flips it, then we re-encode, sometimes the alpha channel is dropped. Tested on a known-bad scan and got wrong colors. TODO: route through a pillow helper that strips profiles first.
• backend/app/workers/cleanup.py:zombie_requeue — there's a race where two workers both see the same zombie and both re-enqueue it. Idempotency key on the task would fix this, but right now we just rely on the dedup window in the dispatcher. Acceptable for now, not for a million tasks/day.
• backend/app/services/result_cache.py — Redis cache for completed tasks. Hit rate is good in the common case (user re-downloads), but the TTL refresh on read is off-by-one. Need to fix and add a metric.
• backend/app/converters/document/docx_converter.py:to_pdf — falls back to LibreOffice when pandoc can't render embedded SVGs. The LO path is ~5x slower. Not worth optimizing until we hit >100 PDF req/min.
• backend/app/api/v1/routes/convert.py — accepts target_format from query string, but a few clients send it in the body. The OpenAPI spec says query only, but we tolerate body. Pick one. (Issue: #14 in my head, not a real one)
• Frontend useTaskProgress reconnects on close but doesn't back off exponentially. Cheap to fix, just haven't.
• Mini-program: progress bar in task-detail is not yet bound to the WS channel — we still poll. WS is wired up in api/client.ts and stores/auth.ts but the task page falls back to the REST endpoint. PR ready in feature/mp-ws.

If you spot a real bug not in this list, please open an issue.

make test # full suite
make test-fast # skip integration
cd frontend && pnpm test

I aim for ~80% on the backend. The converter layer is the lowest coverage because most of the work is shelling out, which a unit test can't really cover without a 200MB LibreOffice fixture.

• JWT signed with RS256. Private key never leaves the API host.
• File uploads: magic-byte sniff before any processing. The antivirus scanner is a stub (no ClamAV in dev) — wire it up via core/antivirus.py.
• API keys use a 32-char random prefix + SHA-256 hash. Constant-time compare in core/security.py.
• All outgoing webhook URLs are SSRF-checked against the private/loopback ranges in core/ssrf.py.
• CORS is locked to the configured origins, not *.

Numbers from make load-smoke on my laptop (16 cores, NVMe, no GPU):

• PDF → DOCX: ~1.8s/10MB
• DOCX → PDF: ~2.4s/10MB
• PNG → WebP: ~40ms/2MB
• CSV → XLSX: ~120ms/100k rows

Workers autoscale by celery_queue_length in K8s; the Helm chart is in deploy/k8s/.

MIT. See LICENSE.

Open an issue. I read them.

— @badhope