I am Yeongjun Yoo, an AI-native builder based in Seoul.
I did not start from a traditional CS path. I started from business, operations, sales, writing, and the very practical need to make messy work run faster. That background shaped how I build software: I care about the user, the workflow, the deal, the deployment, and the feedback loop at the same time.
My current focus is simple:
Build real systems with AI agents, then make them reliable enough for actual work.
I work across agentic coding systems, full-stack product delivery, education technology, and business automation.
|
Claude Code, Codex, MCP, custom skills, multi-agent workflows, eval harnesses, LLM-as-judge review, preference loops. |
Next.js, TypeScript, React, Supabase, Postgres, auth, payments, Vercel, QA, and production handoff. |
|
I like demo-led sales: build a working version early, use it to clarify the real problem, then turn the conversation into delivery. |
Co-founder of Quest-On, exploring how exams, feedback, and classroom thinking should change in the AI era. |
- Built an internal 27-skill, 7-agent Claude Code workbench for sales-to-delivery automation, with evals, rubric scoring, and human-edit feedback loops.
- Co-founded Quest-On, an AI assessment product used in real university classrooms with paid exam runs.
- Worked as a sales-fluent AI engineer/FDE across client delivery: requirements, quoting, full-stack build, QA, deploy, and handoff.
- Led partnership and sponsorship execution as Hongik University Student Welfare Committee Chair, securing KRW 100M+ in festival sponsorships.
- Built a profitable e-commerce operation before university life fully settled in, reaching KRW 25M monthly revenue with zero paid marketing.
- Still writing, reading, competing, and building because the best software usually comes from people who are not only software people.
Most of my client work is private, so here is the public part with numbers attached.
| Repo | What it is | Measured |
|---|---|---|
| onprem-mcp-data | On-prem MCP data platform. A rule-based router sends a question to vector search, NL2SQL, or knowledge-graph traversal, exposed as 7 MCP tools. | Benchmarked 3 embedding models on a 68-question gold set: hit@5 0.985 (bge-m3) vs 0.779 (hash) vs 0.368 (nomic). Knowledge-graph recall 0.278 to 1.000 after fixing 4 defects. |
| korea-agent-production-kit | Korean-first OpenAI Agents SDK reference: one read-only approved-source tool, structured citations, explicit review_required, zero external action tools. |
Kept a false-retrieval failure and turned it into a regression fixture. 5 offline cases, 18 unit tests, 5 live gpt-5-mini runs. |
| queryward | Catches N+1 queries in your tests, before they hit production. | |
| alf-gate | Turns support logs into ALF-compatible test cases and regression-checks resolution rate and policy violations across knowledge changes. |
The private side is where most of it lives: an internal AI delivery workbench (39 skills, 7 agents, 32 shared modules, 50-case eval set), Quest-On classroom deployments, and enterprise and public-sector delivery from discovery to handoff.
1. Find the real workflow, not the polite version in the first brief.
2. Ship a working artifact fast enough to create honest feedback.
3. Add the harness: specs, tests, evals, reviews, logs, and rollback paths.
4. Keep the business loop attached: user, buyer, operator, maintainer.
5. Repeat until the system survives real usage.
I like people who move quickly, write clearly, and care about whether the thing actually works.
I am building toward a career at the intersection of:
- AI engineering
- Forward deployed engineering
- Developer tools and agent runtimes
- Education technology
- Small, durable software businesses
The long-term bet is that a small team with strong taste, strong automation, and strong feedback loops can ship more than a much larger team without them.