Most engineering teams hit a wall when production reliability and HPC/research-grade infrastructure need to coexist. Web-scale distributed systems engineers don't speak HPC. HPC researchers don't ship production AWS code. Healthcare AI teams need both, and almost no one has both.
I've watched the same pattern repeat:
- Legacy systems quietly cap throughput, but nobody redesigns the data path β they tune it
- Research code lives on a laptop until it dies on the cluster
- "We'll add compliance later" becomes "we shipped, now what?"
The cost? 50% over-spending, 30% latency you can't fix with caching, and an LLM you can't actually ship into a regulated environment.
The rare engineer who bridges production AWS at scale, HPC-based ML/AI research, and regulated healthcare LLM deployment β three intersecting depths that almost never coexist in one person.
π₯ Regulated Healthcare LLM (funded RA, Summer 2026 β):
- Clinical LLM companion for older adults with MCI at SLU HPC Lab (research prototype, pilot after IRB approval) β Read More
- LLM design: safety handled outside the model (rule-based crisis check before every LLM call; the AI second pass can only raise risk), prompts assembled from config files the PI can edit, post-session LLM metadata extraction
- AI-driven development: ran a Claude Code multi-agent team with explicit rules, context docs, and CI gates β 184 PRs and ~43k lines across backend, web, and mobile in 5 months; verified agent output myself (e.g., found a crisis sentence both safety layers missed by running real-model scenarios)
π¬ HPC + ML Research at SLU HPC Lab (M.S. CS, Dec 2026):
- Federated Learning parallelization with empirical Amdahl analysis (1.27Γ stable speedup, ~24% parallel fraction ceiling proven); now extending to multi-node FL + FedAsync on the SLU Libra cluster
- Gaussian Process Regression + LSTM pipeline for clinical force prediction with uncertainty estimation, on NVIDIA L40S under SLURM
- HPC-CARLA: closed-loop evaluation of autonomous-driving agents on HPC with GIST (2 nodes Γ 8 A100, SLURM + Singularity) β contributed the watchdog + auto-resubmission job chain; co-authored paper accepted at WORKS 2026 (SC26 workshop)
βοΈ Production AWS at scale (DoveRunner, 2023β2025):
- Redesigned a synchronous-RDB watermark token system into a Lambda + Redis + SQS/SNS architecture β 5M+ tokens issued across Seoul and Oregon, later Frankfurt, with a 30% latency reduction
- Built a self-hosted AWS EKS transcoding pipeline replacing a Hybrik-based SaaS β 50% cost cut, 30% throughput lift; owned GOP-aware segmentation, transcoding workers, and segment merge
- Reduced 100-user video conference room initialization from 0.6s to 0.1s β a 6Γ user-facing latency improvement at TmaxWAPL
Mid-level (SDE II / SWE II) roles in:
- βοΈ Cloud / Distributed Systems Infrastructure
- π€ ML / AI Platform Engineering
- π¬ Video / Streaming Infrastructure
- π₯ Healthcare AI
π St. Louis, MO | Open to relocation anywhere in the U.S. | F-1 OPT + 3-year STEM extension eligible | Available January 2027
β Reach me via LinkedIn or email tofoth@gmail.com.
β Clinical LLM Chatbot for MCI Patient Support β Research prototype (funded RA, Summer 2026 β) π Read More
- iPad-based voice conversation companion for older adults with Mild Cognitive Impairment (MCI), built at SLU HPC Lab as groundwork for a clinical feasibility pilot supporting a future NIH R01 application. Pilot planned after IRB approval; code private (clinical constraints)
- How the LLM is used (design):
- Safety outside the model: every message passes a rule-based crisis check before any LLM call; a crisis signal stops the conversation and starts a lock β safety question β 911/988 β staff alert flow that only a person can clear. A context filter cuts false positives, and the AI second pass can only raise the risk level
- Prompts as config: persona, always-on safety rules, participant context, a topic bank (5 domains, 14 categories), and a 5-minute wrap-up instruction are separate files assembled per request, so the PI can review and edit them without code changes
- LLM after the session: the model extracts topics and clinical-domain flags as research metadata; transcripts and audio are not stored in the database, and token usage is logged per call for the staff dashboard
- Voice latency and cost: AWS Transcribe + Polly with per-sentence TTS and speculative early STT, about 1.3β1.6 s to first audio on dev; Polly engine chosen on latency (~1.0 s vs 2.7β2.9 s) and cost ($16 vs $30β100 per 1M characters)
- How it was built (AI-driven development): a Claude Code multi-agent team (lead, server, frontend, mobile, AI, QA) β 184 PRs, 178 tasks, ~43k lines across backend, web, and mobile in 5 months
- Context design: each instruction is self-contained (background, files and scope, what not to touch, how to verify, PR steps); architecture, DB, and API docs are kept as agent context, and a task is not done until the docs are updated
- Team rules: folder ownership per role; no test means not done; verify on a running server, not only unit tests; QA re-runs instead of trusting "done"; safety and session-limit changes merge only after my review
- Automated gates: CI for backend, web, and mobile on every PR; dev deploys only after green CI; prod only from tags already verified on dev; isolated test environment
- Verifying the agents: ran the PI's 15 crisis scenarios against the real model and found a sentence both the rules and the AI pass missed β added a pattern and a regression scenario; checked a "mobile voice works" report and found zero STT calls β split into a device test; rolled back a breaking deploy instead of patching prod, then diagnosed on a branch and reintroduced it
Stack: FastAPI, React + TypeScript, React Native, PostgreSQL, Claude API, Claude Code, AWS (Transcribe, Polly, EC2), Docker, GitHub Actions
β HPC-CARLA: Closed-Loop Evaluation of Autonomous-Driving Agents on HPC β Paper accepted at WORKS 2026 (SC26 workshop), with GIST AIGS π Read More
- Co-authored (2nd author) a system that runs closed-loop CARLA evaluation of six public agents (TCP, InterFuser, CILRS, NEAT, Roach, LAV) as declarative YAML pipelines over one stage library, across 2 nodes Γ 8 NVIDIA A100 GPUs with SLURM and Singularity/Apptainer
- Contributed the watchdog + automatic-resubmission job chain so long-running evaluation jobs survive crashes and allocation expiry without manual intervention
- System results: persistent per-GPU CARLA servers reuse the ~120s boot across jobs; episode-granular accounting recovered 94% of 13,059 route evaluations from jobs that job-level accounting would have discarded
Stack: CARLA, SLURM, Singularity/Apptainer, NVIDIA A100, Python
β Parallelizing Federated Learning Client Simulation π Read More
- Identified the single-process bottleneck capping FL research throughput; prototyped Python
ProcessPoolExecutorparallelization and diagnosed a PyTorch tensor-pickling deadlock at the IPC boundary, resolving it via NumPy serialization - Measured ~1.27Γ stable speedup across 50β200 Non-IID MNIST clients on an 8-core CPU; proved an Amdahl ceiling (~24% parallel fraction), pointing the next iteration to multi-node FL rather than further single-machine tuning
- In progress: scaling FL to multi-node SLU Libra HPC and Asynchronous FedAvg (FedAsync-style); planned: CIFAR-10 + ResNet-18 workloads
Stack: Python, PyTorch, NumPy, multiprocessing, MNIST, FedAvg
β Dental Aligner ML Force Prediction π Read More
- GPR for thickness extrapolation: Found XGBoost collapsed to flat lines outside the training range (0.25/0.5mm), useless for predicting forces (FxβFz, TxβTz) on U6/U7 molars at unseen thicknesses (0.75/1.0/1.25mm). Re-architected with Gaussian Process Regression (Matern kernel) for ΞΌ Β± 2Ο uncertainty; augmented training with weighted-delta-method synthetic 0.75mm data to distinguish 1.0 vs 1.25mm predictions
- LSTM LOO + Partial Observation forecasting: Reframed as "given a new patient's early measurements, forecast later force/moment trajectory." Designed a 60-experiment matrix (5 cohorts Γ 3 horizons Γ 4 sheets) β train on 4 other cohorts plus target cohort's first (11βk) points, predict the last k β {1,2,3} points
- Model: LSTM(128, 2 layers, dropout 0.2) β Dense(64) β scalar, 300 epochs with StepLR
- Complementary design: GPR covers spatial extrapolation (thickness), LSTM covers temporal extrapolation (14-day window). Trained on NVIDIA L40S GPUs under SLURM on the SLU Libra HPC cluster
Stack: Python, PyTorch, GPyTorch, XGBoost, scikit-learn, pandas, numpy, SLURM
β Crypto Lead-Lag Strategy Validation π Read More
- Tested the day-trading idea "altcoins follow BTC/ETH with a delay" against real market data before building any trading bot. Claude wrote the code; I set the hypotheses, ran the data collection (exchange APIs were reachable only from my machine), checked the results, and made the calls
- Scope: 9 strategies and ~20 variants. Every test used a random-entry control, an in-sample/out-of-sample split, and per-coin measured round-trip cost; each tool was first checked on synthetic data
- Pushing back on the AI changed the results: questioned fixed-time exits β a 4-way exit engine exposed that fixed exits recorded a +1.10% win as -0.20%; questioned the 1:1 BTC-follow assumption β per-pair beta removed fake signals (fewer than 30 left); asked for the missing pure trend-following test β worse than holding (-5.2% vs -2.5%); out-of-sample check turned +27.7% into -31.0%. Rejected AI proposals that did not fit (kimchi-premium retest, futures funding, long order-book collection)
- Result: all 9 strategies rejected β the signal (~+0.1%) never beats the round-trip cost (0.115% to 1.022%). No real trading
Stack: Python, pandas, NumPy, Upbit / Binance / Bithumb APIs, Claude
β Watermark Token Issuance Optimization π Read More
Redesigned and built the synchronous-RDB token service on a non-linear-index Lambda + Redis + SQS/SNS architecture, to fix slow issuance under heavy traffic and cross-region index collisions blocking scale-out, delivering a 30% latency reduction and 5M+ tokens issued across Seoul and Oregon, later extended to Frankfurt.
Stack: Kotlin, AWS Lambda, Redis, SQS, SNS, CloudWatch, RDS
β Distributed Encoding System on AWS EKS π Read More
Built a containerized encoding pipeline on AWS EKS, owning GOP-aware video partitioning, transcoding workers, and segment merge, to replace a Hybrik-based external SaaS that was cost-prohibitive and inelastic at scale, cutting encoding cost 50% and lifting throughput 30%.
Stack: Kotlin, Spring Boot, AWS EKS, Docker, Redis, FFmpeg
- M.S. in Computer Science (in progress) β Saint Louis University, USA, Aug 2025 β Dec 2026 (expected)
Advisor: Prof. Ted Ahn | Lab: High Performance Computing Lab | GPA: 3.67 / 4.0 - B.S. in Computer Engineering β Ajou University, South Korea, 2015β2018