I build backend systems, infrastructure, and AI products - with concurrency, observability, and recovery designed in.
My work connects three things: useful products, explicit system guarantees, and evidence that the implementation holds up. Today, that means a voice trading startup, production AI agents, and contributions to the tools those systems depend on.
- Co-founder & CTO · TrabyOS - Lead engineering across the voice trading product, backend, and native/web integration.
- Software Engineer Intern · Bizarre Cube AI / MXN Commerce Group - Build production AI agents for fashion e-commerce operations. Organization
- Computer Science · Soongsil University, 2022–present - Coursework across 12 subjects.
7 changes accepted across 6 upstream repositories - six directly merged PRs and one change cherry-picked by a maintainer. My focus is observability, cancellation, state consistency, and regression tests. Status checked 2026-10-19.
OpenTelemetry · From one instrumentation bug to shared test infrastructure
- Turn a PR-local test into a maintainer-requested shared utility. During review of my WSGI fix, an OpenTelemetry maintainer asked for a common sampler in the core test utilities. I implemented
CapturingSampler, released in 1.45.0 and now reused in the tests of 8 HTTP instrumentation packages: WSGI, Django, ASGI, aiohttp-server, Falcon, Flask, Pyramid, and Tornado. Core #5681 · Adoption example - Make headers available when sampling happens. Fixed WSGI request-header collection at span creation, with sanitized values kept out of internal spans and metrics. Merged with approvals from three maintainers. Contrib #5032
- Follow the change into the rest of the ecosystem. Verified and reviewed 6 follow-up instrumentation PRs. Found Tornado ignoring sensitive-header masking settings and fixed it using the shared capture helper. Contrib #5170 · Tracking issue #4975
- Ouroboros #2365 · Cancellation cleanup - merged, v0.54.5. Fixed leaked HTTP resources and stale connected state when MCP initialization is cancelled; added 6 regression tests.
- psycopg #1345 · Async pool cancellation - reproduction & verification. Reproduced swallowed cancellation, supplied a deterministic regression test, and verified the maintainer's fix, released in psycopg_pool 3.3.2. This contribution is separate from the 7 accepted changes above.
- Micrometer #7925 · Tested observability documentation - merged. Documented custom Java HttpClient observation tags with examples compiled and exercised by the test suite.
- Ponytail #828 · Branch-specific state restoration - merged, v4.13.0. Restored the selected branch's saved mode after conversation-tree navigation.
- Caveman #1013 · Reliable forced updates - adopted in v2.7.0. Repaired rule-file updates for 3 editors with regression tests. The maintainer cherry-picked the change into #1015.
- 🏆 Manifest Special Award · UNITHON 2026 · marketvalley
- 🥈 Silver Prize · Soongsil University CS Software Competition 2026 · Cham Domi
- 🏅 Excellence Award · Soongsil University Solved Code Algorithm Competition 2025
TrabyOS · A spoken request becomes a reviewed trade
Co-founder & CTO. A voice trading workspace for web and macOS: ask for market context, review an order, revise it, and approve execution on your exchange account. The product also provides voice technology for partner platforms.
- Own the path across languages and surfaces. Python orchestration, Spring Boot services, native Swift UI, and the web experience connect market data, account state, and order workflows.
- Separate interpretation from execution authority. Natural language prepares the action; explicit user approval and deterministic checks govern execution. Market identity, quantities, and account state stay part of the transaction contract.
- Build beyond the conversation. Exchange integration, real-time updates, onboarding, packaging, and deployment make the system usable outside a development session.
Python · Java / Spring Boot · Swift / SwiftUI · TypeScript · PostgreSQL / WebSocket
Official website · GitHub organization
ssuAI · ssuMCP · ssuAgent · ssu-ai-service
Students search university information and use their own academic, LMS, and library data through a web app or an agent. I built the four-service platform, from connectors and orchestration to deployment and observability.
- Campus workflows through both a web app and an MCP server. Expose 52 MCP tools for university information and academic, LMS, and library tasks, usable from clients such as Claude Desktop and Cursor. REST and MCP share business rules; LangGraph manages specialist routing, persistent conversations, and approval before writes. MCP connection guide
- 100 competing reservations → 1 upstream write. An intent queue and same-seat suppression protect the university system. Verified with 100 concurrent requests against WireMock; this is a reduction in external writes, with a deliberate latency trade-off. Load-test evidence
- Built and operated the delivery path. ARM64 images move through GitHub Actions, GHCR, and ArgoCD to k3s. Prometheus/Grafana, Tempo, and Loki connect metrics, traces, and logs; fault injection verified circuit-breaker isolation and recovery. Observability & incident records
Java / Spring Boot · Python / LangGraph · PostgreSQL / Redis · k3s / ArgoCD · OpenTelemetry
marketvalley · One hypothesis, a complete market-validation workflow
Backend, AI, and platform. A founder describes a problem and solution once; the team product creates a landing page, five social cards, copy, and a Meta ad, then brings back Insights, visits, and consented early-access reservations.
Built the durable workflow behind it: PostgreSQL jobs, SKIP LOCKED, leases, operation keys, and checkpoints keep generation, ad creation, and result collection independent of a browser tab, with controlled retries around external side effects. Architecture
🏆 Manifest Special Award · UNITHON 2026.
Next.js · Supabase / PostgreSQL · Anthropic · Meta Marketing API
Organization · Source & team roles · Live service
Geuneul · Find relief from the heat, without waiting on the map
A summer survival map over 150k+ public points of interest, with nearby shade, shelters, and live community reports. I built the data ingestion, spatial backend, application, and infrastructure.
- Found the bottleneck in the execution plan. Spatial indexes were already working; expired reports were still scanned for every scored search. Indexing their expiry filter cut SQL execution time 361 → 172 ms (~52%) on the same synthetic dataset: 300k places and 210k reports, measured locally under QEMU. EXPLAIN comparison
- Kept data and updates repeatable. Idempotent ETL, PostGIS radius/nearest-neighbor queries, SSE reports, and infrastructure as code connect ingestion to the map.
Spring Boot · PostgreSQL / PostGIS · SSE · Terraform / AWS
Source & architecture · Open the map
Cham Domi · From dormitory eligibility to an actual roommate group
Frontend, roommate domain, integration, and application infrastructure. Built explainable recommendations across eight lifestyle dimensions, recruitment/application flows, and one-to-one/group chat. Owned the application's Spring Boot/MySQL deployment, Helm configuration, network isolation, image-digest checks, and backup/restore gates.
🥈 Silver Prize · Soongsil University CS Software Competition 2026. The team delivered the dormitory eligibility and roommate experience together.
Spring Boot · Next.js · MySQL · WebSocket · Helm
Organization · Open the service
HeungMap · Turn a forecast into a decision a planner can use
Backend, data, modeling, and integration. A tourism-data product that connects festival planning with visitor discovery. The model forecasts regional visitor-days, with D−30 information constraints, prediction ranges, and contributing factors.
- WAPE 4.256% → 3.939%. A seasonal baseline plus LightGBM residual model improved WAPE by 7.43% relative on a retrospective time holdout of 4,959 region-days. Model card
- Forecast coverage 30 → 98 out of 100 real events. Corrected date-window and administrative-region history handling to make more events usable without changing the prediction target to festival attendance. Decision record
- Keep generated reports grounded. The API rejects unsupported figures; scheduled retraining only adopts models that pass five acceptance gates.
FastAPI · LightGBM / pandas · Next.js · Claude · Scheduled data pipelines
Folding · Ask across documents; approve the changes
Agent backend and file-operation boundaries. A local-first workspace for HWP, Office, and PDF. Built the FastAPI sidecar, agent orchestration, approvals, file edits/undo, and knowledge synchronization. Retrieval exposes source evidence; approved edits produce and verify a new file while preserving the original.
FastAPI · Electron · SQLite · Local LLM
Organization · Product repository · Agent backend · Run locally
Trip Sheet · JECT · Give a team a repeatable deployment path
Team infrastructure. Built the infrastructure foundation for a collaborative product: three OCI ARM servers separating production application, production data, and development/monitoring.
- Build once, promote the same image. Reusable GitHub Actions workflows build ARM64 images, deploy to development, and support manual promotion to production. Environment-level deployment coordination and health-check rollback protect the delivery path.
- Include the work after deployment. OpenTofu, Docker Compose, Nginx, PostgreSQL/Redis, backups, and Prometheus/Loki/Grafana cover provisioning through operations. Deployment and timeout behavior were checked in an isolated integration environment; application rollout is being connected.
OCI / OpenTofu · GitHub Actions / GHCR · Docker / Nginx · PostgreSQL / Redis · Prometheus / Grafana / Loki
JECT · Infrastructure · Backend · Frontend
| Work | What I worked on |
|---|---|
| Macro · UNITHON24 | Voice access to existing kiosk software. Reworked fragile coordinate automation into UIA/OCR actions, postcondition checks, and an idempotent SQLite queue; 76 recorded tests and Linux/Windows CI. |
| AI말평 experiments | Compared 19 encoder backbones / 24 configurations with nested out-of-fold evaluation; further checked finalists with leave-one-prompt-out evaluation and leakage audits under deployment constraints. |
| AXWAR submissions | Built three pipelines for grounded FAQ answers, disclosure summaries, and refund calculation, separating model-generated text from deterministic numeric and evidence validation. |
| Area | Stack |
|---|---|
| Backend & data | Java · Spring Boot · Python · FastAPI · PostgreSQL/PostGIS · MySQL · Redis |
| Infrastructure & observability | Docker · Kubernetes/k3s · ArgoCD · GitHub Actions · Terraform/OpenTofu · AWS/OCI · OpenTelemetry · Prometheus/Grafana · Tempo/Loki |
| AI & product integration | LangGraph · MCP · RAG · LightGBM · TypeScript/Next.js · Swift/SwiftUI |
Performance figures link to their measurement conditions. Project roles distinguish individual ownership from team delivery.



