AI Service Engineer — document intelligence, retrieval systems, and reliable product delivery

Email · GitHub

I turn ambiguous documents and operational rules into software that can explain where its answer came from. I have 5+ years of experience building and operating software, and I currently focus on production AI systems.

• Document intelligence — PDF parsing, OCR/vision, structured extraction, rule engines, and human-review states
• Retrieval systems — hybrid search, reranking, source-page traceability, guardrails, and evaluation
• Reliable delivery — FastAPI services, databases, caching, CI/CD, packaging, monitoring, and runbooks

Built an end-to-end workflow for complex appraisal PDFs: parsing and OCR, document structure recovery, bank-specific checks, numeric-role validation, four-state review results, persistence, alerts, packaging, and QA.

Python PyMuPDF Naver CLOVA OCR Gemini / Vision FastAPI MSSQL pytest

Source code and client data are not public. The portfolio describes architecture and engineering decisions without exposing confidential material.

Designed a retrieval workflow that combines dense and keyword search, reranking, question routing, terminology normalization, source-page links, hallucination guards, cache prewarming, and a repeatable evaluation set.

FastAPI BGE-M3 Qdrant SQLite FTS5 Redis Jina / BGE reranker Docker Compose RAGAS

A Korean–Japanese–English travel interpreter for situation phrases, text input, and two-way live voice. The implementation includes 16 kHz audio resampling, 100 ms streaming chunks, TTS caching with device fallback, circuit breaking, secret management, CI/CD, rollback, and operational runbooks.

FastAPI Cloud Run PWA Android WebView iOS WKWebView

A tennis-club operations product covering balanced draw generation, members and roles, events, scores, statistics, social features, PWA, and native packaging.

My contribution visible in the public Git history:

• 949e628 — resolved a Capacitor 8 / Apple Sign-In SwiftPM dependency conflict using a persistent patch-package patch; restored the production app icon; added iOS export-compliance metadata and related native build changes
• 298015b — advanced the iOS marketing version and build number for the subsequent App Store Connect upload

Attribution note: Tenol's product scope, architecture, and repository-wide output are team outcomes. I do not present the full application as my individual implementation.

1. Make the failure modes and operational constraints explicit.
2. Keep the original evidence reachable from every generated answer.
3. Separate model judgment from deterministic validation.
4. Ship evaluation, rollback, and observability with the feature.