Deterministic LLM Localization Pipeline (TradeTEST-ES)

Showcase report: 01_live-dashboards/tradetest-es-bilingual-qa-report.html · Code: 02_ai-engineering-code/tradetest-es-pipeline/

Problem. Translate ~3,400 content items (DB rows, quiz questions, lessons, UI strings, affiliate pages) from English to Mexican Spanish — unattended, without quality drift.

What I built. A resumable CLI pipeline (Bash orchestration + Python helpers) wrapping OpenAI Codex through a 5-stage quality chain:

1. Build a style guide + term glossary

2. Translate

3. QA audit for errors

4. Back-translate to English to catch meaning drift

5. Final confidence gate (0–100 score, LOW/MED/HIGH risk)

Engineering details that matter.

Idempotent & resumable — every step writes a file; re-running skips finished work.
Fault tolerant — 5-minute timeout per call + one automatic retry on hang.
Safe writes — insert.sh shows a dry-run preview before any DB write.
Reviewable output — per-batch HTML report with side-by-side EN/ES, risk, confidence, and drift items.
Cost-aware — ~$0.50 smallest batch, ~$60–150 full run; spend controlled by batching.

Why it matters. This is the core applied-AI competency: turning a non-deterministic model into a deterministic, auditable, fail-closed pipeline with QA gates and cost discipline.

GG · Applied AI Engineering portfolio · synced to youdontneedmy.help design