Backend engineer focused on turning AI features into systems that survive real traffic — not just working demos.
- I also have hands-on experience across Java/Spring Boot and SvelteKit-based product development.
I care about three things after implementation: performance, reliability, and operational cost. I also use AI agents as part of the engineering workflow itself — delegating repeatable implementation and verification work while keeping architecture, security boundaries, and acceptance decisions under explicit human ownership.
- Reduced unnecessary LLM call volume and token usage through synthesis-based batching and lightweight context payloads
- Reduced OOM risk by moving large-file processing out of long-lived JVM heap residency and into temporary-file based sequential processing
- Improved a search API from 24.04s P95 to 2.014s, then verified the root cause with
performance_schema,EXPLAIN, and k6 instead of stopping at the first plausible hypothesis - Built and documented an AI-agent orchestration workflow with SPECs, acceptance criteria, review loops, and human approval gates
- Languages & Frameworks: Java 21, Spring Boot, Spring Data JPA, QueryDSL, SvelteKit, TypeScript
- Databases: MySQL, PostgreSQL, pgvector
- AI / Agentic Development: LLM APIs, LangChain4j, Claude Code, MoAI-ADK, prompt / agent workflow design
- Performance & Observability: k6, Prometheus, Grafana, Micrometer, HikariCP, MySQL
performance_schema - Infra & DevOps: Docker, GitHub Actions, Vercel
🧠 Cubrain — AI-powered PDF study SaaS
- Reduced unnecessary LLM call volume and token usage by combining related pages into synthesis-based batches and sending lightweight coordinate/text payloads
- Added exponential-backoff retry for
429 Rate Limitfailures to improve service reliability under external API constraints - Reduced JVM heap pressure and OOM risk by storing multipart uploads as temporary files, processing them sequentially with PDFBox, and deleting them after completion
- Structured LLM input around the user's selected PDF region and surrounding context so source context is preserved instead of relying on prompt changes alone
⚡ sbb_board — Backend performance engineering
- Reduced a 1:N lookup from 6,506ms → 70ms (98.9%) using semi-join style queries and HashMap-based DTO assembly
- Under VU100 load, improved search P95 from 24.04s → 2.014s, reaching 40.5 req/s, 0.00% HTTP failures, and a 0.44% 3s SLO exceed rate
- Tested the HikariCP saturation hypothesis by changing pool size from 20 → 40 → 80 and discarded it when the exceed rate stayed around 50–52%
- Traced the real bottleneck to an
ORcondition preventing FULLTEXT index use, then rewrote the query withUNION + INNER JOINand verified result equivalence at 1,475 rows
🎙️ EchoBloom — AI-powered affirmation speaking service (team of 6)
- Normalized STT output whitespace and implemented Levenshtein-distance similarity scoring with an
O(MN)dynamic-programming approach - Used an 80% similarity threshold with up to 3 retries to turn noisy STT output into a more explainable pass/fail flow
- Shared API boundaries and voice-processing flows as diagrams so frontend and backend members could discuss edge cases against the same structure
🛡️ secure-file-upload — AI-agent orchestration + server-side upload validation
- Defined requirements and acceptance criteria first, then used MoAI-ADK + Claude Code to delegate implementation and verification work
- Kept human approval around architecture, security policy, secret handling, and other high-risk decisions
- Recorded 46 decision markers in
PROMPT_LOG.md: 24 accepted directly, 17 accepted after revision, 3 rejected, 2 deferred - Of the 17 revised decisions, 11 were corrected by the AI verification loop itself and 6 required direct human intervention
- The repository is intended as supporting evidence of how I decide what to delegate to AI and what not to delegate
- Proposed a SPEC-review guardrail that separates AI analysis from human decision authority
- Introduced ideas such as “Analysis is Pull, Not Push” and “Detect → Explain → Ask, but never decide” to reduce anchoring and keep product decisions under explicit human control
- The maintainer confirmed that the proposal identified gaps in the current pipeline, added it to the backlog, and selected the anchoring-prevention approach for a priority experiment
Domain, Harness & Agent-Agnostic
An experimental control plane for agentic software engineering.
Exploring how workflow state, memory, policy, and execution context can live above individual agents, harnesses, and providers.
- Google Gemini 3 Hackathon — Lucidify: Built a multimodal lucid-dreaming interface with SvelteKit and GCP serverless tooling within the hackathon window
- AI 포텐데이 x Naver Cloud Hackathon: Implemented STT normalization and Levenshtein-based similarity scoring for EchoBloom, and shared API/service-flow diagrams during a 6-person team project completed within the 20-day hackathon period
Paper-first notes on self-improving AI agents
Studying inference-time scaling, verification, evaluation, planning, memory, and authority-aware agent architectures.
Research ideas → system design → practical agent architecture.
More compute improves capability.
It does not automatically grant more authority.
Linear Algebra → Three.js → WebXR
Learning 3D mathematics twice: first through Mathematics for Machine Learning, then again by making each concept visible with Three.js.
“I wanted JARVIS🤖. Unfortunately, I had to learn linear algebra first.”
📫 How to reach me:



