AI Engineer building agentic, multi-agent tooling and the data systems underneath it. MISM (Business Intelligence & Data Analytics, AI Management concentration) at Carnegie Mellon University, Aug 2026 to Dec 2027.

Most recently the sole engineer on a production AI expense platform at Accuver America (Python, Flask, SQLite, Docker): receipt OCR, statement matching, 729 merged PRs, a 5,100+ test suite behind CI gates, and a Gemini-to-Claude fallback chain benchmarked against a 90-receipt golden set.

• LLM / AI: RAG, prompt caching, structured output, multi-agent orchestration, LLM evaluation and golden-set benchmarking, LangChain, Anthropic / OpenAI / Gemini APIs, MCP, ChromaDB, OCR extraction, Text-to-SQL
• Data & Engineering: Python (Flask, pytest, pandas, scikit-learn), SQL (SQLite, MySQL, DuckDB, PostgreSQL / Supabase), R (caret, glmnet), Java, Docker, GitHub Actions, Playwright, Streamlit, React, Power BI (DAX), process mining (ProDiscovery)
• Certifications: AWS Certified AI Practitioner, AWS Certified Cloud Practitioner, Lean Six Sigma Yellow Belt

• AI Expense Platform (Accuver America, 2026): receipt OCR, statement parsing, receipt-to-statement matching and exception flagging that cut reconciliation time 90%+; benchmarked 7 LLMs, cut LLM cost ~90% with prompt caching, shipped nine schema migrations on the live database with rehearsed drills
• Weekly finance tooling (Accuver America): rolling sales forecast automation (about 6 hours of weekly manual work to a one-minute run), a DuckDB + Streamlit revenue-forecast comparison tool with a guarded Text-to-SQL assistant, and a Salesforce pipeline change tracker
• 5G-domain RAG prototype (Accuver America): LangChain + ChromaDB + Claude over 7 standards whitepapers (~700 chunks) with a Streamlit UI
• Sustainability data audit (Andersen Corporation, 2025): Python + Power BI data-integrity audit across 50+ sites and 20+ vendors that surfaced $20K+ in duplicate billing and weighing errors
• Process mining (PuzzleData, 2024): 3M+ funnel log rows on a life-insurance engagement whose recommendations lifted final-stage conversion 5%; led a procurement proof of concept (1,200+ events, ~80-85% conformance)

• Sharday: collaborative 2-4 person travel planner (React 19, Supabase with RLS, Edge Functions, realtime sync, Cloudflare Pages); 251 commits, 19 migrations, used by 5 people
• Walmart Purchase Prediction: regression model comparison in R (linear, ridge, lasso, RF, GBM) on 550K+ transactions; log-linear best at R² 0.74
• Process Mining Portfolio: procure-to-pay event log from three systems, discovery and conformance checking (~80-85%)
• 보다 (boda): vision OCR (Gemini to Claude) that turns receipt images into structured expense data and Excel

Seven open-source plugins for Claude Code, used throughout the 729-PR Accuver codebase and adopted by the successor engineer:

Parallel workflow

• ddaro: worktree-based parallel workflow with deletion-aware staging, crash-recoverable context, and CI-orchestrated merges

Multi-model review

• prism: multi-angle code review, 5 parallel agents plus a singleton verifier pass
• triad: 3-perspective deliberation (clarity, longevity, comprehension) until consensus
• mangchi: cross-model code hardening, Claude writes while Codex CLI critiques one axis at a time
• prism-devil: single-agent attacker-mindset review with auto-loaded checklists

Deep research as verified RAG

• batchim (받침): verification-gated research with an isolated verifier, 3-panel entailment checks, code-enforced span/number anchors, and a sha256-signed run ledger

Receipts

• boda (보다): the vision OCR pipeline above, packaged as a plugin

Technique cards (not counted above)

• galmuri (갈무리) distills reusable techniques from a project into cards; knowledge is the library it writes to

I treat AI coding as an engineering system, not a single chat:

• Parallel, isolated branches: worktrees keep concurrent work from colliding (ddaro)
• Cross-model review gates: Claude and Codex critique each other before anything merges (prism, triad, mangchi)
• Verification before trust: research and retrieval must cite, anchor, and pass entailment checks, not just sound right (batchim)
• Distill what works: recurring solutions become reusable technique cards (galmuri to knowledge)

• Email: minwoo.park219@gmail.com
• LinkedIn: linkedin.com/in/mp74484
• Portfolio: minwoopark.dev