01
AI capability
Connect natural language and model output to product behavior.
Chat To Do · Todaktodak
I build and operateAI services people actually use.
I connect AI capabilities to product flows, then carry them through reliable operations and reusable knowledge.
Explore the work
About Me
AI Service Engineer
Semin Kim
I go beyond model output, turning it into services that teams and users can depend on.
See the work02
Service integration
Make frontend, backend, and AI work as one user journey.
PinLog · AlgoSu
03
Delivery & operations
Observe resources, failures, and logs; design safe delivery paths.
PinLog · AlgoSu
04
Knowledge reuse
Preserve project decisions and outcomes as verifiable evidence.
Graph RAG Second Brain
Selected Work
SELECTED WORK
From AI capability to operable services
Five projects trace a path from natural-language features and generative AI to team operations, solo products, and reusable knowledge.
TEAM AI OPERATIONSTeam of 6Infrastructure & DevOps Lead
PinLog
Operated a team AI service under tight resource constraints.
A place journal that preserves why a place mattered and retrieves it through natural language.
k3scontainerdArgo CDPrometheusFastAPI
Single VM / 4 vCPU / 15 GiB
Team AI Platform
DELIVER
GitHub Actions · Argo CD
RUN
k3s · containerd
OBSERVE
Sentinel · Human decision
PinLog API → Atomic Queue → GPU Worker → Result
Problem
Frontend, backend, AI, and observability had to share one 4 vCPU / 15 GiB VM.
Contribution
Identified Docker and cri-dockerd as CPU bottlenecks, moved to containerd, and built Helm/ApplicationSet GitOps and observability.
Result
Standardized delivery, caught a real sync failure, and connected a separate GPU server through a safe asynchronous queue.
KEY IMPLEMENTATION
Optimize
Measured CPU throttling → containerd
Deliver
Immutable images, infra PRs, Argo CD
Observe
Sentinel summarizes evidence; people act
Async AI
Atomic claims, requeue, bounded failure
VERIFIED SIGNALS
1.44%containerd CPU
1.10Max CPU PSI avg60
6/6Cowork team adoption
AI search quality and keyword fusion were advisory work and are not claimed as direct implementation.
AI PRODUCTSolo projectSolo product, engineering & operations
AlgoSu
Built and operated AI code review as a real service.
An AI review service that explains complexity, readability, and improvements beyond correctness.
FastAPINext.jsAI Agentsk3sObservability
SUBMISSION PIPELINE
Code in, actionable review out
ANALYZE
Complexity and approach
REVIEW
Role-based AI review
SYNTHESIZE
Grounded improvements
RetryStateObservability
Problem
The challenge was not one AI response, but operating the whole review flow through failures and retries.
Contribution
Separated review stages, designed async jobs, retries, and state tracking, then deployed and operated it on OCI.
Result
Extended a solo AI project from build and deploy into ongoing, evidence-driven operations.
KEY IMPLEMENTATION
Review Pipeline
Role-based reviews and synthesis
Resilience
Async state, retries, visible failures
Operations
Prometheus, Grafana, Loki
OPERATING PRINCIPLE
Deployment is the beginning of operations, not the end.
KNOWLEDGE OPERATIONSPersonal systemSolo design, engineering & operations
Graph RAG Second Brain
Operate experience and decisions as evidence-backed knowledge.
A local knowledge system linking documents and conversations as claims and evidence, reusing only human-approved facts.
Graph RAGSQLiteEmbeddingsObsidianHuman-in-the-loop
EVIDENCE GRAPH
Source → Claim → Evidence → Answer
Source
Claim
Evidence
Verified answer
When evidence is insufficient, the system abstains.
Problem
As documents grew, current facts mixed with stale claims and provenance became difficult to verify.
Contribution
Separated sources, entities, claims, and evidence while preserving review status and provenance paths.
Result
Generates portfolio and application material from verified evidence and abstains when support is insufficient.
KEY IMPLEMENTATION
Ground
Claims and evidence linked to sources
Review
Reuse only human-reviewed facts
Regress
Retrieval and answer regression checks
OPERATING PRINCIPLE
Choose a verifiable abstention over an unsupported answer.
AI SERVICE FOUNDATIONS
The AI foundations that led to service operations
Chat To Do connected language to product actions; Todaktodak combined multiple generative models into one service flow.
GENERATIVE AI PIPELINE
Todaktodak
Connected multiple generative AI models into one user journey.
An emotional support service generating a four-panel comic, music, and sentiment analysis from a journal entry.
MY ROLE
AI resources & pipeline contributor
WHAT IT TAUGHT ME
Learned that product flow and collaboration matter as much as individual models.
Generative AIAI PipelineBackend Collaboration
NATURAL LANGUAGE TO ACTION
Chat To Do
Turned natural language into executable product actions.
A conversational task service connecting user intent to concrete product functions.
MY ROLE
AI resources & pipeline contributor
WHAT IT TAUGHT ME
Moved beyond answer generation to systems where user language triggers real product behavior.
Function CallingRAGIntent Routing
HOW I WORK
Build it, operate it, make it reusable
Across projects, the method stays the same: start from the user problem and preserve operational evidence.
01
User problem
Define where users get stuck before choosing technology.
02
AI capability
Connect model output to concrete product behavior.
03
Operational proof
Observe failures, resources, and logs; improve from evidence.
04
Knowledge reuse
Preserve decisions and outcomes as verifiable knowledge.
Automation handles repetition; people retain responsibility for operational decisions.
Blog
Blog
Writing
Documenting lessons learned from practice.
Contact
Learn. Build. Grow.
Open to hiring inquiries, tech discussions, and collaboration.