Deokhwan Kim

Resourceful problem solver · Product engineer

[email protected]

github.com/IISweetHeartII

log8.kr

What I can take on

I turn ambiguous customer problems into working AI products and reliable delivery systems.

Product discovery, customer interviews, MVP scope, and full-stack implementation.
Cloud infrastructure, deployment, observability, recovery, and code-review systems.
Agent orchestration with explicit ownership, validation gates, and human review boundaries.

Selected evidence

AgentGram: 881 organization-wide merged pull requests; 789 were merged by the autonomous pipeline I designed and operate. The figure measures system throughput, with authorship distributed across autonomous work and review.
Fleet operations: 45 scheduled jobs and 383 completed kanban cards.
log8.kr: 61 hand-written blog posts, 100 agent-published Library posts, and 2 guides.
GitHub stars: agentgram 32 and ax-score 3.

Projects

PurplePepper — zero-downtime production infrastructure and unmanned observability

2026.03 — Present | Backend & AI Engineer

Ran controlled experiments to prove the overnight batch bottleneck was memory, not CPU; cut processing time (Takt Time) by 41%.
Found and fixed a backup pipeline silently mis-reporting "success" daily (recovered 232GB of internal infra backups, not customer data); folded the fix into the shared template to block the same failure elsewhere.
Audited 505 AI agent sessions to compute failure rates and installed hooks blocking risky commands; an autonomous agent fleet checks 408 crawling servers daily and reports anomalies.

Migrated Supabase to NestJS + AWS (ECS Fargate) with Terraform-based blue-green deployment.
Zero-downtime legacy migration (39 caller accounts unchanged, one-line rollback, 38K requests/day).
Converted the overnight batch platform (8 boxes, 64 lanes) to OS-level self-healing — 490K+ jobs/night at 98.4% success, including Mac/Linux fleet + AWS setup.

Shipped the Nyangmoney reward app on iOS — designed a B2B2C user-acquisition channel

2026 — Present | Product & engineering (iOS shipped, Android closed beta)

Built an independent reward app to solve the company's B2B2C cold start (needing advertisers and users at once) via a dedicated acquisition channel.
Designed an ad-consuming reward loop so a user pool forms before advertisers are signed, reversing the usual sequencing.
Shipped to the iOS App Store; resolved native social-login issues on real devices while running Android closed beta.

Next.js + Capacitor single codebase across web/iOS/Android.
Supabase + Firebase (FCM, Admin) for auth/push/data, with AdMob ad-reward integration.
Native Google/Apple social login and an asset pipeline for app store review.

Built GCP infrastructure and established team code review as Backend Lead

2025.12 — Present | Backend Lead (5-person backend team)

Raised review frequency 400% (about 5×).
Reduced Terraform-based recovery from 56h to 5h.
Scripted the rebuild path for an estimated about 2h next lead time.

Java 21, Spring Boot 3.4, MySQL, QueryDSL.
Terraform + GCP infrastructure as code.
CI/CD, VPC, separated dev/prod databases, rollback deployment.

Designed and operates AgentGram's autonomous contribution pipeline

Ongoing | Pipeline designer and operator

The Hermes multi-agent pipeline autonomously implements, reviews, and merges 789 of 881 organization-wide pull requests; branch protection and CI gates keep autonomous merges safe.
Kanban-based autonomous decision-making decomposes the backlog and assigns tasks without human intervention — only irreversible calls escalate to a human approval gate.
Review-gated autonomous work across a multi-repository ecosystem.

Next.js 16, TailwindCSS v4.
Supabase (PostgreSQL + RLS), API-first architecture across core/JS/Python/MCP/AX repos with Ed25519-signed identity.
Self-hosting, Docker, REST API, OpenAPI.

Won the ICAN-LABs top award with a 40-interview MVP

2025.09 — 2025.12 | Solo full-stack developer, frontend-led

Cut mathematics-content production time to 1/10.
Identified tool complexity through 40 interviews.
Used a 5-week I-Corps-based validation sprint.

Next.js, MathJSON, Fabric.js, JSXGraph.
Formula parsing and geometry visualization engine.
MVP prototyping and user testing.

Tech Stack

LanguagesTypeScript, Java, Python, C++

FrontendNext.js, Astro, React, TailwindCSS

BackendSpring Boot, NestJS, Supabase

DatabasePostgreSQL, MySQL, pgvector

DevOpsAWS (ECS Fargate, ECR), GCP, Terraform, Docker, Cloudflare Tunnel, GitHub Actions, Sentry, tmux

AI Orchestration Lab

Runs Claude Code, Codex, and agy (Gemini) concurrently via OMC/OMX-driven Loop Engineering (generate → test → verify) to ship fully verified work.
Measured 5× learning efficiency and 10× development speed after the workflow shift (vs. January 2026).
Hermes fleet: runs 24/7 on a single Mac mini with 45 scheduled jobs and 383 completed kanban cards, atomic claim → isolated workspace → written handoff.
Autonomous work areas: automated blog publishing, GA data analysis, YouTube video production and upload, and AgentGram's autonomous development and QA.
Irreversible work stops at a human approval gate, so autonomy and safety hold together.

Experience

UMC 9th Cohort — Web Part Lead (teaching and mentoring), Spring Boot challenger

2025.09 — Present

UMC 8th Cohort — Web challenger (React, Next.js)

2025.03 — 2025.08

Practical AI lectures — senior small-business owners

2025.05, 07

Practical AI lecture — Re:ium founders (MCP, agent AI)

2025.09

Awards & Certifications

Awards

Certifications

Education

Chung-Ang University2020.03 — 2026.08

Mathematics (primary major) / Software & Humanities Convergence (double major).
Completed the teaching curriculum (Level 2 Secondary School Teacher Certificate).
162 credits completed; GPA 3.86/4.5.