Sung Jun "Tony" Baek

LLM AgentOps Engineer

I specialize in LLM AgentOps—turning agentic prototypes into observable, testable, cost-aware, and reliable production systems. I also build manufacturing planning platforms where complex operational workflows must become validated and explainable decisions.

Agent Harnessing & Workflow Engineering

I design Claude and Codex agent harnesses that turn open-ended development tasks into controlled, repeatable, and reviewable engineering workflows.

Define scoped context and task boundaries so each agent receives the minimum information and authority required for its role.
Configure tool permissions and controlled execution paths for repository, browser, document, and delivery workflows.
Orchestrate plan → implementation → test → review → PR workflows with explicit checkpoints and recovery paths.
Use verification gates, reviewable artifacts, and traceable handoffs to make outcomes reproducible and auditable.
Separate implementation and review roles when independent validation improves confidence.

Symphony — LLM AgentOps for APS Analytics

A production AI-agent platform where LLM AgentOps practices make natural-language analytical workflows observable, validated, cost-aware, and reliable for advanced planning and scheduling.

Built and refined multi-step agent behavior across planning, SQL-assisted analysis, validation, replanning, and report generation.
Centralized LLM execution paths so runtime context, token usage, inference traces, and cost signals remain consistent across agent workflows.
Added regression coverage for prompt injection, execution context, and SQL tooling to protect behavior during refactoring.
Expanded production observability with Prometheus metrics and Grafana dashboards across agent, API, database, model, session, and tool layers.
Contributed operational controls for usage reporting and protected API documentation access.

Taelim — Manufacturing APS & Scheduling Engine

A manufacturing planning engine for order allocation, machine scheduling, and lot composition in corrugated-packaging operations.

Prevented stale demand from being processed twice during lot composition.
Fixed confirmed-order recomposition flows so eligible orders remain in factory-specific planning.
Refactored coating domain rules and strengthened work-in-progress and order-length data integrity.
Added repeatable QA coverage with unit tests, integration datasets, and GitHub Actions workflows.
Improved operational support with Oracle schema/index versioning, configurable logging, and engine-result exports.

Additional Platform Experience

Financial Services Platform — Real-Time Fund Processing

Software Developer at BeaconFire Inc., contributing to a real-time fund-processing platform for financial services.

Strengthened availability in OpenShift/Kubernetes by resolving liveness-probe issues, reducing recurring health-check failures from 40 per hour to zero.
Designed multi-process, event-driven services with Kafka and AMQ/RabbitMQ across Azure SQL, PostgreSQL, and DynamoDB, and refactored legacy code into an order management system.
Optimized database-intensive batch processing for 100,000+ orders through caching, batching, and indexing.
Supported Azure disaster recovery and failover, Sumo Logic/Log4j2 monitoring, and Veracode vulnerability remediation.
Established Docker/Flyway test environments and applied TDD with JUnit 5 and Mockito to improve deployment stability.

Warehouse Digital Twin Service Platform

Full-Stack Engineer at VisionSpace, building a hybrid-cloud digital-twin service platform for warehouse operations.

Designed Java/Spring Boot microservices across on-premise and AWS infrastructure, with Jenkins delivery pipelines targeting EC2, ECR, and ECS.
Built a React/TypeScript frontend and web application server for visualizing and interacting with dynamic digital-twin simulations.
Developed a .NET/C# stress-testing application that simulated high-concurrency robotic-device traffic through MQTT, WebSocket, and AWS IoT.
Integrated LLM, RAG, and computer-vision capabilities with the platform to expand intelligent search, analysis, and visual processing workflows.
Supported production connectivity and scaling with Route 53, ACM, S3, security controls, and load-balanced AWS services.

Technology focus

LLM AgentOps	Backend & event systems
Claude · Codex · agent harnessing · workflow orchestration · context engineering · verification gates · LLM workflows · runtime traces · token and cost signals	Python · FastAPI · Java · Spring Boot · Kafka · RabbitMQ · PostgreSQL

Observability & cloud delivery	Digital twin & manufacturing
Prometheus · Grafana · GitHub Actions · Jenkins · Docker · Kubernetes/OpenShift · AWS · Azure	MQTT · WebSocket · AWS IoT · React · TypeScript · Oracle · APS scheduling

Earlier experience

ModelTranslator — model-based IoT tooling that translates UPPAAL models into Python-oriented embedded-system workflows.
Hibernate quiz web app — a Spring Boot and Hibernate web application for working with relational data and protected routes.
D* algorithm — pathfinding exploration implemented with Python/Pygame and Java desktop UI tooling.