Joshua (Sung Hyun) Kim SungHyunC

My projects span machine learning pipelines, AI agent workflows, and practical web applications.
I enjoy turning a problem into a working tool, then making its behavior and limitations easy to understand.

_EDUCATION Korea University, Sejong Campus B.S. Computer Convergence Software	_GRADUATING February 2027
_{LOOKING FOR} Software engineering and applied AI roles New grad · United States	_{WORK AUTHORIZATION} U.S. citizen

Awards

5th Place, Indie Developer Track ($25,000)
Meta Global AI Developer Hackathon, November 2026; 5,367 projects were submitted.
Project: Video2URDF, built solo.

Excellence Award
Industry-Linked Capstone Design Competition, Korea University Sejong Campus, spring 2026.
Project: Multi-Agent BEMS.

Research

Undergraduate Research Assistant, Korea University · Sep 2025 – Present
Multi-agent reinforcement learning for air combat

Reimplemented JSBSim F-16/A-4 flight dynamics and an air combat environment in JAX, cutting per-policy training from an estimated 190 CPU hours to ~20 GPU hours; verified statistical equivalence to JSBSim in 5 stages.
Reproduced the main results of an IEEE ICA 2025 hierarchical MARL paper over 10 seeds (62 policies, 1.1M evaluation engagements), tracing remaining gaps to seed variance and undocumented ammo and altitude rules.
Extended an F-16 simulator to 2v2 combat with 7 RL setups, 131 automated tests, and Unreal Engine 5 rendering.

Python JAX JSBSim Unreal Engine 5
_{The code is private (lab research).}

Selected projects

A laptop filmed with a phone, with the measured model drawn over it: the lid in yellow and the base in purple

Video2URDF

5th Place, Indie Developer Track, Meta Global AI Developer Hackathon. A Muse Spark agent turns one phone or iPhone LiDAR video of an articulated object into a measured URDF/MJCF using SAM 3.1 part tracking and least-squares joint fitting, then checks it by replaying the model in MuJoCo over the video.

Python SAM 3.1 Muse Spark MuJoCo

Repository · Live demo · Demo video

Schematic of a one-on-one engagement: the agent's aircraft closing on its opponent, which sits inside the cone ahead of the agent's nose

F-16 Air Combat AI Agent

AI Pilot Top Gun Challenge · Summer 2026

Trained a PPO dogfighting policy on a 6-DOF F-16 flight model and wrapped it in rule-based safety layers, logging zero crashes or dropped responses across thousands of local and dozens of live-server rounds.
Built a competition-rules simulator, unattended server orchestration, and 83- and 122-opponent evaluation pools; every release had to pass league, held-out, live-server, and bit-identical checks.
Traced aim overshoot in live-server logs (0.4 s on target vs. 1.0–1.4 s for a reference model), then added an arbitration layer that raised head-on score from 0.39 to 0.88 over 64 paired rounds.

Python Ray RLlib JSBSim
_{The code is private: the repository includes the competition's environment files.}

The MA-BEMS dashboard: a 3D campus with live KPIs, the agent decision log and the six agents' status

MA-BEMS: Multi-Agent Building Energy Management

Capstone project, fall 2026. Six agents (coordinator, HVAC, lighting, PV, ESS, EV charging) run a simulated campus side by side with a rule-based BMS on identical inputs.

Cut energy cost by 16.7 ± 0.2% vs. the rule-based BMS and held peak demand under the 1,000 kW limit (0 vs. 1,996 minutes over) in simulation; ablations and nine stress scenarios, including one where the agents do worse, are reported.
Fine-tuned Qwen3-8B with QLoRA on simulator-generated data for the local AI assistant: answers passing a number check rose from 86.4% to 98.3%, and proposed actions that would change nothing fell from 121/228 to 3/209.
Deployed as a static dashboard on GitHub Pages, with an access-keyed LLM gateway behind a Cloudflare tunnel.

Python JavaScript Three.js QLoRA Ollama

Repository · Live dashboard

The Multi-Agent BEMS dashboard: energy KPIs, a power chart, and responses from the monitoring, analysis and report agents

Multi-Agent BEMS (spring 2026)

The spring 2026 capstone that preceded MA-BEMS (Excellence Award): a LangGraph supervisor routes operator questions to monitoring, analysis and report agents over simulated building-energy data, with rule-based alerts and fallbacks.

Python LangGraph FastAPI Next.js

Repository

Observed and CNN-predicted remaining useful life over flight cycles for two test engines

Engine RUL Prediction

Parallel HDF5 processing and CNN experiments on NASA N-CMAPSS simulated engine data, with recorded results and evaluation caveats.

Python PyTorch HDF5

Repository · Report (PDF)

ETF Answer Agent

A prototype question-answering workflow with retrieval, validation gates, and revision routing.

Python LangGraph

Repository · Live demo

Team contributions

MOA — AI-powered civic reporting: Contributed AI classification, the React admin console, and report grouping and prioritization in a university industry-practice team project. See the merged classifier and admin console pull requests. The linked deployment currently demonstrates the frontend; the backend runs locally.
BidderBidder iOS: Contributed login/signup flows, Firebase phone verification, search UI, and SwiftLint CI through merged pull requests for a used-goods auction app.

Tools used in these projects

Python JavaScript Swift SQL

PyTorch JAX Ray RLlib scikit-learn JSBSim MuJoCo Three.js Ollama Unreal Engine 5 LangGraph FastAPI Streamlit React Next.js Git

For implementation details, setup instructions, and experiment evidence, start with the linked project READMEs.

Navigation Menu