Jui Ting C.

Python Backend Developer | LLM API, Quant, Financial Pipelines

Need AI integrated into a system that already works? I build production Claude/LLM backends — Tool Use, prompt caching, structured output, streaming — and the data pipelines they sit on top of. Python that ships on time, ships with tests, and runs in production, not scripts that break next week. What I build with AI: → Claude/LLM integrations — Tool Use orchestration, prompt caching (51.8% hit rate measured in production, not estimated), structured output, SSE streaming, FastAPI endpoints. Production-grade scope enforcement and cost control. → AI features inside existing apps — added Claude-powered exception detection to a client's IFRS 9 accounting app without touching the frozen calculation engine. The hard part is not the API call, it is knowing what not to break. → Workflow automation — n8n integration (cron + REST dispatch + status polling), FastAPI inference services with automated retrain workflows, Telegram and Discord bots with persistent state and error recovery. Recent work: → IFRS 9 loan accounting app in Streamlit — code hardening, UI conversion, and a Claude-powered exception detection feature. The client, a non-engineer, wrote: "Excellent at explaining technical work in plain language, even without a technical background." 5.0 stars. → Bitfinex P2P lending bot in production — multi-key architecture, watchdog with automatic recovery, 138+ hours across deployment, debugging, and daily monitoring. Running unattended. → SEC EDGAR 10-K parser for Business Development Companies — 103 companies, $1,324,753K reconciled to the audited Total Investments line, 58/58 tests passing, Docker. Also: → Quant & trading systems — signal generation, multi-timeframe funnels (ADX, EMA, Bollinger, ATR-based sizing), backtesting with look-ahead detection, exchange APIs (Binance, Bitfinex, Deribit), ClickHouse and SQLite tick storage. → Financial data engineering — SEC EDGAR filings, multi-source ETL across financial APIs, Pandas + PostgreSQL pipelines, idempotent ingestion with deduplication and audit logging. → ML & classification pipelines — fastText, scikit-learn, PyTorch feature extraction, confidence calibration. How I work: every project Dockerized, deployed on DigitalOcean (or your cloud), version-controlled, tested, documented. I check my own work before you have to — on one research project I found a look-ahead bias in my own backtest and reported it rather than shipping the number. I respond within hours, ask scoping questions before quoting, and ship on time. If something breaks after delivery, I fix it. Bilingual EN/中文 native, Taiwan-based. Available 30+ hours per week. Comfortable with US morning hours.