Portfolio

Portfolio

The visual portfolio is the gallery; this page is the evidence map. Each thread connects the problem, what I built, and evidence readers can inspect. Private data and unpublished work are clearly separated from public evidence.

Evidence boundary. Shipped internal systems, public repositories, and published technical analysis are different evidence types.

Project map

Project threadProblem, contribution, and evidence
Recent side projectsProblem: Apply evidence-aware workflows to distinct creative, game, and discovery problems.
Built: Five independent projects developed alongside a shared agent-harness and skills practice.
Evidence: Project details and the visual portfolio.
AI products & RAGProblem: Make private studio knowledge searchable, grounded, and operable.
Built: Retrieval pipelines, agent workflows, product surfaces, and operational controls.
Evidence: Project details, retrieval writing, and the visual portfolio.
Agent harnessProblem: Help tool-using agents finish bounded work without hiding failure.
Built: Execution loops, edit integrity, critic gates, reusable skills, and verification contracts.
Evidence: Public repositories and agent writing.
Multimodal QAProblem: Turn gameplay screenshots and video into reviewable QA signals.
Built: CV, OCR, detection, VLM interpretation, and report workflows.
Evidence: Research details and multimodal writing.
Game AI & automationProblem: Make game evaluation and production workflows reproducible.
Built: Simulation agents, procedural systems, Unity tooling, and agent-driven builds.
Evidence: Project details and playable projects below.

Recent side projects: one development practice, distinct products

These five separate projects share a development-system context: jeo and jeopi explore spec-first planning and verification harnesses, while jeo-skills packages reusable skill contracts. That shared practice does not mean every runtime or provider is connected, or that every project is production-verified. The visual portfolio includes their images and captured demonstrations.

FrameForge Studio

A local/admin video-production hub carrying source-backed research through versioned storyboards, layout, optional generation, rendering, and human review. Its bounded workflows include version and paid-generation guards, output hashing, and approval tied to the reviewed output. The visual portfolio shows fixture-data UI captures; no public deployment is claimed.

Ttoktak Brain (Echo Turn)

A portrait Unity puzzle where P follows the current command and Q repeats the previous accepted command; A/B plates open the exit. I built the turn-based puzzle with undo, restart, and optional guidance. Its visual portfolio image is a Unity Editor capture, not evidence of a device build.

Wonder Scanner

A browser camera collection game that runs object recognition locally and turns everyday objects into 80 collectible Wonders. The scan, resonance, and capture loop includes a codex, album, quests, and a mobile-friendly fallback. Open Wonder Scanner; the visual portfolio’s UI sequence does not show a live camera scan.

TRACE-RPG — The Sealed Lighthouse

A Godot turn-based investigation micro-RPG where generated events are proposals and deterministic symbolic commit gates decide canonical state. The work includes parsing, validation, bounded repair, and hash-linked receipts. The dashboard and repository provide public engineering evidence, not player efficacy or usability results.

Notchima (놓치마)

A Korea-wide family-event discovery service with map search, five-axis filters, and source-backed listings. Its file-first pipeline ingests public APIs and web search, normalizes events, and builds JSON and SQLite indexes for a static UI. Explore the app and repository.

AI products: retrieval in real workflows

SAGA RAG search engine operations interface SAGA connects retrieval quality to a product surface that a game studio can actually operate.

At Supercent, I built AI products where retrieval was only the middle of the system. SAGA indexes 594 game-design documents and 1,563 vectors across 15 games, combining RAG-Fusion, CRAG, hybrid search, cross-encoder reranking, and a grounding-oriented generation pattern. Millie extends that direction into per-user knowledge operations and a Slack-driven local-agent system. Brain focuses on citation-first knowledge operations with incremental ingestion and connected editing workflows.

I extended retrieval into operational workflows: cited answers, incremental knowledge updates, connected sources, and agent-assisted actions. I see this progression from RAG search to a system teams can operate as groundwork for productizing AI agents, rather than a claim that retrieval alone makes an autonomous product.

The portfolio shows these as shipped products. The related writing explains the reusable engineering questions:

The practical lesson is that retrieval quality, permissions, observability, and user experience cannot be reviewed independently. A good answer is not enough if nobody can trace its source, correct the knowledge, or understand why the system chose a tool.

Agent tooling: verification is part of the product

jeo-code coding-agent harness jeo-code is one part of a public tool family built around bounded execution and verification.

My open-source agent work is public enough to inspect directly:

  • jeo-code is a Bun-based coding-agent harness with a multi-provider tool loop, anchored-edit integrity, and a verify-before-done contract.
  • jeopi explores a critic-gated plan, execute, and verify spine.
  • jeo-skills packages reusable workflows across agent runtimes.
  • oh-my-jeo provides a spec-first workflow layer around chat agents.
  • jeo-claw connects Discord commands to bounded repository work.

The associated articles are not release announcements. They document design pressure and failure boundaries:

I treat agent orchestration as earned complexity. One capable agent with clear tools and a visible completion contract is usually better than a large cast of agents with vague authority.

Multimodal QA: evidence, not verdict

AutoQA tooling running against a game screen AutoQA work combines classical vision, learned detectors, and language-based reporting rather than assuming one model can own the whole verdict.

My graduate research and personal tooling connect image-based QA, OCR, multi-scale template matching, fine-tuned detection models, video context, and VLM-assisted bug reporting. The interesting problem is not whether a model can describe a frame. It is whether a team can turn visual observations into a reproducible QA workflow with known false negatives and reviewable evidence.

This thread includes an IEEE RAAI 2024 poster on image-based game QA automation and a 2025 publication on automated QA reporting from natural-language captions. The multimodal, video and game AI section of Start Here provides the wider technical context behind that work.

Internal data and unpublished research artifacts are not public evidence. Where I cannot share a dataset, customer document, or source repository, I state the boundary instead of replacing it with a stronger claim.

Game AI and automation: repeatable production

Castle War physics siege gameplay Castle War is a playable project, but the main experiment was the coding-agent-driven production workflow behind it.

At NCSOFT and Com2uS, I worked on simulation-based difficulty evaluation, cellular-automata and GAN-based generation, reinforcement-learning prototypes, Unity-to-Python systems, and internal production tools. That experience shaped a simple bias: an AI feature is unfinished until the surrounding workflow is measurable and reproducible.

Recent public work carries that idea into build automation and agent-driven game production:

The games matter, but so does the method: explicit build contracts, generated assets with review boundaries, repeatable verification, and honest reporting when the environment cannot complete a native build.

How I make the writing inspectable

Search rankings, GitHub stars, and a polished screenshot are weak substitutes for evidence. My editorial contract is more specific:

  1. Who: Every article is published under my name and links back to About, where my production and research background is explicit.
  2. How: Repository articles start from primary sources. When I claim code runs, I execute it or label the limitation. Version-sensitive findings are dated.
  3. Why: The goal is to help engineers make a decision, reproduce a result, or avoid a failure mode. It is not to summarize every trending repository.
  4. Corrections: Readers can report a mistake through Contact. I verify concrete corrections and update the article rather than silently preserving a stronger claim.

This does not make every post equally useful to every reader. It does make the origin, method, and boundary of the work visible.

Further reading

The visual portfolio records what I built; this page and the technical writing expose the reasoning, evidence, and open questions behind that work.

© Jang Young Jeong. Some rights reserved.

Search article titles, categories, and tags. Full text is searched when needed.

Type to search articles.