I'm an architect, and I built Architectena, a professional site for architects (portfolio, CV, architecture-only jobs, office pages and a studio workspace) that runs fully in Arabic and English. It's Django 5.2 on PostgreSQL, with Celery for background work.
Translating the interface was the easy part. These four problems were not in any translation file, and each one looked fine until it didn't.
1. The year 2026 rendered as "2,026"
With USE_THOUSAND_SEPARATOR = True, Django groups every integer a template renders. It can't tell a quantity from an identifier. A project year came out as 2,026, and the same rule turns {{ obj.id }} into 1,234 inside form values and URLs.
We searched the templates and found 323 places that render an identifier. 214 of them would actually break: 189 form values that would submit 1,234 and fail validation, and 25 links that would 404. None had broken yet only because no table had passed 1,000 rows. A year was simply the first integer to cross the line.
The fix was one setting, not 214 |unlocalize filters:
# settings USE_THOUSAND_SEPARATOR = False Grouping isn't lost where you want it: {{ n|intcomma }} forces grouping, so follower counts still read 1,234. Use it on quantities, never on ids or years. A per-template filter would also have missed every template written later, which is how the year sites drifted in the first place.
2. The second catalogue
The .po catalogue looked perfect: 13,031 entries, zero untranslated, zero fuzzy. But Arabic pages still showed English filter bars.
The gap was the other catalogue: the name_ar columns on vocabulary models (skills, specialties, office fields, job fields). 164 of 309 values were blank. Nothing rendered empty, because every model's display_name() fell back to English. So the symptom wasn't a missing string. It was an Arabic page with English chips on it, which reads as a site that was localised by somebody who stopped halfway.
Two lessons:
- Audit data, not just strings. A catalogue checker can't see database columns.
- One mapping for shared terms. "Landscape architecture" appears as an office field, a specialty and a job field. Two Arabic spellings of one term is how a filter bar starts looking machine-made, so all three tables read from one English→Arabic module.
One deliberate exception: brand software stays Latin. An Arabic CV says AutoCAD, not a transliteration, so those rows keep an empty Arabic name and fall through to English on purpose. The module names them explicitly, so the next coverage report doesn't file it as a bug.
3. Arabic number formats are a separate setting
Django looks up {module}.{lang}.formats for locale formatting. A project-level override lets Arabic pages group numbers the way the house style wants without touching English:
FORMAT_MODULE_PATH = ["locale_formats"] # locale_formats/ar/formats.py Small, but it's the difference between numbers that look right in a right-to-left sentence and numbers that look pasted in.
4. "No AI training" without delisting yourself from search
The site promises users that their work is not training data. A promise is just a sentence unless something enforces it, so robots.txt is generated from one Python list of AI-training user agents (GPTBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot and others).
Two details matter:
- A crawler that finds its own name obeys only that group and ignores
User-agent: *entirely. So every agent gets its own full block, not an addition to the shared one:
for agent in AI_TRAINING_AGENTS: out.append(f"\nUser-agent: {agent}\nDisallow: /\n") - The search halves are different bots.
Googlebotis notGoogle-Extended,Applebotis notApplebot-Extended,OAI-SearchBotis notGPTBot. Block the wrong one and you quietly delist yourself while claiming to protect users. A second list of search agents exists only so a test can assert the two lists never overlap.
The public promise page and robots.txt read the same list, so they can't drift apart: a name added to the promise is refused in the same commit, and a test fails if it isn't.
The pattern behind all four
Every one of these passed a normal check: the catalogue was complete, the pages rendered, robots.txt existed. What caught them was asking what the user actually sees in the second language, and then writing a guard test so it can't come back.
If you're building something bilingual, especially right-to-left, I'd love to hear which of these you've hit, or what I've missed.
Architectena is free for individual architects, at architectena.com.
