Byeonghyeon Kim

CAREER PORTFOLIO
AI Engineer · LLM Systems
On-Prem LLM Infra · Access-Controlled Memory · Security-grounded
Gyeonggi-do, South Korea | bk@bkan.dev | (+82) 10-4676-5884 | GitHub | Portfolio | Threads
Career Summary
PeriodOrganization / TypeRole · Scope
2026 Enterprise AI Systems (Freelance) AI Engineer · Technical Lead - on-prem LLM infrastructure · enterprise AI memory · database design · ADR governance
2025 - Present Solopreneur Founder / Engineer - planned, built, and launched AI products (OurFolio · ASHD v2 · BluePy · BK-HQ)
2025.06 - 2026.01 FastCampus AI Bootcamp (14th cohort) Completed - 5 applied ML competitions · MLOps · multi-agent project
2021 - 2025 Franchise Business Owner/Operator Owner/Operator - POS-driven demand forecasting and inventory optimization; prepared for AI career transition
2018.02 - 2021.02 CMT EYES Security Monitoring Analyst - SOC / CERT incident response
Enterprise AI Systems - Freelance 2026
Designed and built an end-to-end AI environment for an enterprise that could not send data to external clouds, spanning LLM infrastructure, data models, enterprise knowledge management, and development processes. Also led technical governance for a small engineering team.
On-Prem LLM Multi-Agent Infrastructure
Role Sole owner for architecture, implementation, and operations  ·  Tech Self-Hosted LLM · Tailscale · GGUF/MLX · Multi-Agent · Discord Bot
  • ContextSecurity requirements prohibited external cloud LLM APIs, requiring fully internal model-serving infrastructure
  • ActionAssigned workloads across 15 Mac Studios, connected them through an isolated Tailscale network, self-hosted LLMs without public exposure, and provided a Discord bot as the single entry point
  • ActionDesigned a local LLM portfolio per device, assigning GLM and Qwen models using quantified license, RAM, quantization (GGUF/MLX), and KV-cache headroom constraints; cross-checked official power data and corrected errors in the hardware sheet
  • ActionBuilt a 2-agent handoff pipeline in which execution and review agents automatically exchanged work
  • ResultDelivered an always-on cluster with automatic restart recovery; enterprise memory and agent systems ran on this stack
Integrated Call-Center and Customer Database Schema
Role Sole owner for data modeling  ·  Tech PostgreSQL · SQL-first Migration (goose) · ETL
  • ActionAnalyzed legacy ETL and designed migration for 94K customers, 86K addresses, and 5.7K consultations; identified IVR/inbound duplication in 32K raw CDR rows and normalized them into 23K actual calls
  • ActionDesigned a 27-table normalized schema integrating call-center and customer data, including natural-key strategy, soft deletion, slowly changing dimensions (SCD), and explicit invariants
  • ActionEvaluated migration engines (rejected Atlas, dbmate, sqitch, and Flyway; selected goose), externally validated pgvector/halfvec capacity limits, and established a SQL-first SSOT plus a schema-invariant checklist as documentation-level TDD
  • ResultPassed 2 external reviews and adversarial code review
Engineering Governance · Technical Lead
Role Primary engineering administrator · decision-record owner  ·  Tech GitHub Monorepo · ADR/MADR · Git Workflow
  • ActionSet up the GitHub organization, monorepo, and branching strategy (main/dev/feat) for a 3-person team; introduced and maintained an ADR (technical) + Decision (non-technical) record system (ADR-001-008)
  • ActionDesigned work allocation across sales, advertising, and CRM tracks; led onboarding; and established worktree isolation and Safe Staging rules for multi-session AI collaboration
  • ResultDelivered a transition PR covering 202 files / +37,000 lines plus a separate summary for a nontechnical team lead; documented infrastructure-related data loss neutrally and proposed recurrence-prevention rules
Byeonghyeon Kim CAREER PORTFOLIO
Enterprise AI Systems - Freelance (Continued) 2026
Team and Enterprise AI Memory System
Role Open-source customization and system architecture  ·  Tech Hindsight · Graphify · Access Control
  • ContextAgent memory accumulated personal work records, requiring both privacy isolation and complete retention
  • ActionCustomized open-source agent memory Hindsight with per-employee personal and team memory banks; controlled Recall and Reflect permissions by security tier so personal banks remained visible only to their owners
  • ActionMeasured 211 production-like records in a 1-week trial, quantified recall-based retrieval limits (15.6% recall), adopted direct time-filtered queries, and disabled the Reflect summary layer to prevent information contamination
  • ActionBuilt an automated synthesis pipeline for daily and weekly reports; measured quality with evals (0% hallucination rate) and applied executive-value curation rules
  • ResultDelivered an enterprise knowledge-management foundation that retained records without violating access boundaries, with a 3-stage personal → team → enterprise promotion model
Korean Embedding and Reranker 3-Way A/B Experiment
Role Sole owner for experiment design and execution  ·  Tech MTEB-ko · FAISS · sentence-transformers · statistical testing
  • ContextA 1.38-point gap on the Korean embedding leaderboard was within noise, so benchmarks alone could not justify replacing the RAG model
  • ActionDesigned an arctic-ko vs KURE-v1 vs BGE-m3-ko 3-way A/B experiment with a pre-specified decision rule (Bonferroni α=0.025, MRR), 90 queries, 6 strategies, and self-retrieval integrity checks
  • ResultNo pair reached statistical significance, and the sample size was insufficient; reported the study as underpowered and retained the model due to no evidence for replacement. Validated the decision with a three-way cross-check: leaderboard results, community signals, and internal experiments
Codebase Knowledge Graph
Role Sole owner for implementation and validation  ·  Tech Graphify · Knowledge Graph
  • ActionStructured the codebase as a knowledge graph with 1,158 nodes / 1,797 relationships / 73 communities and integrated it with team automation tools
  • ActionAudited the full pipeline and found and fixed 4 graph-integrity bugs: hyperedge merge contamination, phantom cross-project edges, noisy-node pruning, and stale state; then finalized metrics
Unified Admin Dashboard (Demo)
Role Sole developer  ·  Tech Spec-Driven Dev (PRD/SDD/TDD) · Accessibility
  • ActionBuilt an admin system deriving 13 modules from one module registry, including accessibility, dark mode, and responsive layouts
  • ActionAudited all flows, identified 12 UX/accessibility gaps, and shipped fixes for 10, including unified search, WAI-ARIA keyboard navigation, dark-mode chart synchronization, and self-hosted fonts
  • ResultUsed spec-driven development to reach 81 passing tests, up from 54, with 0 regressions and 0 console errors across 13 routes
Korean Document Pipeline · DOCX Track-Changes Library
Role Sole developer  ·  Tech Python · OOXML(ECMA-376) · pandoc · MinerU VLM · typst
  • ContextExecutive reporting relied on HWP/DOCX/PDF, but no tool could ingest tracked changes and comments into the LLM pipeline
  • ActionBuilt a pure-XML library for DOCX tracked changes/comments, removing the LibreOffice dependency and using pandoc round trips as the TDD baseline through v1-v4; 48 tests
  • ActionValidated MinerU VLM on Korean PDFs with typst, then packaged the workflow as a Claude Code skill
  • ResultA/B pilot measurements showed a 55% reduction in task time (67 sec → 31 sec) and half as many tool calls
Additional Work
  • ActionConducted full-stack research for meeting recording → diarization → minutes automation (4-cycle deep dive); built and delivered a financial-data validation MCP server (6 tools), explicitly labeling unverified results; maintained the internal web platform
Byeonghyeon Kim CAREER PORTFOLIO
Solopreneur - AI Products 2025 - Present
Owned the full lifecycle from product planning and development through deployment and operating decisions. Launched 2+ product MVPs.
OurFolio - Evidence-Based Portfolio Platform GitHub ↗ ourfolio.bkan.dev/demo · Public Showcase
Role Solo product development (planning, engineering, deployment)  ·  Tech Next.js 16 · Neon · Drizzle ORM
  • ActionBuilt a platform for developers to substantiate work with evidence, featuring server-rendered public profiles, one-click GitHub repository import, and dynamic OG/SEO generation
  • ActionImplemented a 2-stage, zero-downtime expand-contract schema migration, test-enforced Sentry PII scrubbing, and policyVersion-based consent history; operated 7 CI checks spanning Lighthouse, coverage, and migrations
  • ResultIndependently built and deployed the MVP through a validated v1. After assessing market viability, ended monetization and converted it into a public showcase with a read-only live demo, demonstrating disciplined product sunset decisions
ASHD v2 - OCR+LLM Receipt and Warranty Management GitHub ↗ ashd.bkan.dev · MVP Launch → Public Snapshot
Role Solo developer  ·  Tech Next.js 15 · Supabase (Auth/RLS) · OCR · LLM · CI
  • ActionUsed OCR plus rule-based extraction and an LLM to structure receipts and warranties and send email/Telegram expiration alerts. Operated a production-grade service through payment-integration experiments and automated solo operations with 15 CI workflows including evals, automerge, and P0 alerts
  • ResultManaged extraction accuracy quantitatively on the KORIE benchmark (314 examples): field-level F1 improved for store 9% → 100%, purchase date 89% → 99%, and amount 98% → 100%; operated an eval regression gate that blocked PRs below seeded gold-set thresholds
BluePy 2.0 - Open-Source Infrastructure Security Scanner GitHub ↗ v0.1 alpha · MIT
Role Sole developer (ground-up rebuild)  ·  Tech Python 3.12 · Clean Architecture · CLI
  • ContextA 2017 KISA legacy security scanner included validators that returned false PASS results, making its output untrustworthy
  • ActionRedesigned it in Python 3.12 with Clean Architecture; exposed only checks with validated causality by default (Linux permissions; macOS SIP, FileVault, and automatic updates), removed false-PASS validators, and isolated unverified legacy rules as experimental
  • ActionBuilt an asyncssh-based remote SSH scanning architecture for Linux/macOS with PASS, FAIL, and MANUAL outcomes; added a real-SSH Docker E2E fixture and 546 tests
  • ResultReleased a CLI-first scanner with JSON/table output and strict SSH host-key verification as a transparent v0.1 open-source project, translating security operations experience into code-reliability standards
BK-HQ - Solo AI Company Operating System In active development and operation
Role Architecture and operations  ·  Tech Multi-Agent Orchestration · Automation Hooks
  • ActionBuilt an orchestration platform enabling one CEO to operate 15 AI agents across 4 teams (product, legal, marketing, engineering), with 31 skills, 19 automation hooks, and 5 crew workflows
  • ResultIntegrated 5 Smart Skill Chains, social automation, Agent Dashboard, and a compound knowledge system (25+ domains, 100+ documents) into one operating system, validating a solo operating model
Byeonghyeon Kim CAREER PORTFOLIO
Security Monitoring & CERT - CMT EYES 2018.02 - 2021.02
Security Monitoring & Incident Response
Role Security Monitoring Analyst (SOC/CERT)  ·  Tech FortiGate/UTM · FortiSIEM · WAF/IPS · Wireshark
  • ScopeRemotely monitored and analyzed 20+ Fortinet-focused WAF/IPS/UTM/SIEM appliances for 50+ clients, including night/weekend shift rotation and the escalation process
  • ActionHandled 1,000+ events per month across the full CERT lifecycle: detection → log/packet analysis → blocking → policy improvement → reporting
  • ActionTuned SIEM correlation rules and custom IPS/WAF signatures; developed new detection scripts covering diverse security events
  • ActionAnalyzed intrusions end to end - reconstructed timelines from logs and packets, identified attack paths, and wrote incident reports
  • ActionRan emergency DDoS response (range blocking, traffic analysis) and triaged KISA/government threat advisories into client-impact assessments and urgent guidance
  • ActionParticipated in penetration-testing and compromise-analysis projects for major enterprise clients; ran phishing simulation exercises
  • ResultImproved false/true-positive classification accuracy by 50%+ through policy and detection-rule tuning
  • ResultAutomated repetitive reports with scripts, reducing per-event reporting time from 30 min → 5 min
Security Engineering & Client Support
  • ActionOnboarded new clients into monitoring - initial appliance policy setup, baseline tuning, and firewall policy cleanup and migration
  • ActionEngineered client security appliances and operated Linux/Unix servers
  • ActionSupported an on-site 10G network expansion for a major education client whose mock-exam traffic surges caused recurring annual outages - hands-on network installation work
  • ActionSupported the shift to SSL-VPN-based remote-work security during COVID-19
  • ActionAuthored monthly and quarterly security reports and delivered regular client briefings
  • ActionLed new-hire training and owned the onboarding curriculum
ML Competitions & Projects - FastCampus AI Bootcamp (14th Cohort) 2025.06 - 2026.01
Completed competition-driven applied ML work. Metrics below show improvement over each competition baseline.
Toss Ads CTR Built a CTR model on large-scale ad logs; improved LogLoss/AP through sequence features, downsampling, and tuning; placed in the Top 10%
Science QA IR Improved RAG retrieval and ranking with Solar embeddings, rank-graph refactoring, and prompt tuning; MAP/MRR +107%
Document Images Classified 17 types of scanned documents; backbone experiments, TTA, augmentation, and post-processing improved F1 +470%
Dialogue Summarization Summarized everyday Korean dialogue; QLoRA fine-tuning, augmentation, and post-processing improved ROUGE +30%
Real-Estate Prediction Improved RMSE by 58%
Marketing AI Multi-Agent System
  • ActionOrchestrated 6 agents to generate trend, copy, segment, review, competitor, and strategy reports, replacing repetitive research
Movie Rating Prediction MLOps
  • ActionBuilt a TMDB API-based rating prediction service with an E2E automated data collection → training → deployment → monitoring pipeline (CI/CD)
LLM-Based RAG Training Content
  • ActionCreated training materials for practitioners adopting RAG for internal document search, covering BM25, vector, and hybrid comparisons, query expansion, and hands-on Hit@k/MRR evaluation
Portfolio Evidence
Project evidence, including screenshots, code, and live demos, is available at portfolio.bkan.dev. See the 1-page resume at resume-en.html.