Resume
Production AI Systems Engineer · 4 yrs 5 mos AI Infrastructure · Agent Systems · LLMOps · Production ML
Contact · lijihoon1206@gmail.com
Summary
I design, deploy and operate AI systems that actually run in production — from sensor signal pipelines to LLM agents. I decide technical choices on measurable evidence, and I treat finding the failures after shipping as half the job. My master’s research (graph neural networks for behavior classification) maps directly onto my current work (IMU behavior analysis).
Selected Impact
- 4.4× throughput · 3.3× latency on a production inference pipeline (96 → 423 win/sec)
- Designed, shipped and operate a natural-language data agent used company-wide — 4 surfaces: Slack · MCP · REST · app API
- Designed a fail-closed de-identification architecture — outbound calls to external LLMs are blocked unless masking has been applied
- Chose an LLM architecture by A/B measurement — discarded the approach that regressed, and published the cost of the one I adopted
- 1 SCI(E) journal paper · 1 IMU behavior-classification paper in preparation (90% feature reduction, under 1pp accuracy loss) → Research
Skills
| Area | Detail |
|---|---|
| AI / LLM | Claude API (tool-use), LangGraph, RAG (pgvector), MCP, eval design, de-identification |
| ML | LightGBM, PyTorch, MLflow, time-series & IMU signal processing, feature-extraction optimization |
| Language / Data | Python, TypeScript, SQL · PostgreSQL, MySQL, ClickHouse, OpenSearch, Parquet |
| Backend / Web | FastAPI, REST API · SvelteKit 5, Drizzle ORM |
| Infra | AWS (IoT·S3·RDS·Lambda·ECR·OpenSearch), Docker, Cloudflare Access, Linux, cron operations |
Experience
Bodit Inc. — Data Analysis Part, Engineering
2024.12 – present · Livestock IMU behavior analysis (6-axis 25Hz → ML pipeline → farmer app)
Responsible for both analysis models and the internal data/AI infrastructure. The systems below were designed, built, deployed and operated solo.
Internal LLM agent platform — an assistant that answers questions about internal data in natural language. Started as an MVP, moved through agentic workflow to a tool-calling agent.
- RAG knowledge base over 3-tier data routing (MySQL / PostgreSQL / ClickHouse with fallback) — 730 schema documents (95 tables · 635 columns) plus an auto-promotion loop from real query logs
- Decided the architecture shift by A/B measurement. A single-label LLM router regressed accuracy (0.86 → 0.69) and was discarded. Exposing each capability as a tool matched baseline while passing 5/5 compound queries — impossible under the old structure — and asked clarifying questions 5/6 vs 0/6. I recorded the adopted approach’s −5pp single-query cost alongside its benefit
- Fail-closed de-identification — transmission to an external LLM is blocked outright when masking has not been applied. Row-level pre-masking plus a regex backstop covers identifiers that column context alone would miss
- Surfaces: remote MCP gateway (SSO) · active anomaly detection (periodic scan + automated root-cause → alert) · Slack bot · REST API
Behavior classification inference pipeline — 4.4× throughput — reworked the production core package that extracts 903 features from 6-axis IMU raw signal and classifies behavior.
- Replaced the bottleneck third-party library with a NumPy reimplementation. I verified numerical equivalence against the existing output before swapping, so the bottleneck went away without changing results
- Selectively vectorized the slowest feature set — 61 of 63; the remaining 2 stayed per-row because that was the right call
- Moved model management from a local-file fallback to a single MLflow registry path
| Metric | Before | After | Gain |
|---|---|---|---|
| Throughput | 96 win/sec | 423 win/sec | 4.4× |
| E2E pipeline | 65.8 s | 19.9 s | 3.3× |
Measured on 24h of real-animal raw data — 2.16M rows / 8,640 windows · full suite 208 passed / 0 failed
Company-wide knowledge base — privacy-safe compilation pipeline — internal conversations compiled and maintained by an LLM. Answers come from compiled documents, not from retrieval over raw text.
- Pipeline: 12-channel ingest → two de-identification layers (regex + GPU NER) → LLM compilation → wiki → agent queries
- Tuned de-identification precision in both directions — tightened rules so identifiers never leave in plaintext and institution names became one-way hashes; loosened rules that over-masked public information, and documented both
- API key failover — switches to a spare key on balance exhaustion, but only on
403 + billing_error - Final review and merge stay human — NER accuracy is under 95%, so the last gate on personal data is not automated
Also — real-time livestock accident detection (IMU anomaly, shadow deployment) · business management system (requirements to deploy in 4 days, 1,090 tests, session layer built on the standard library alone) · analytics dashboard (SvelteKit 5, Role × Group ACL + audit log) · edge inference feasibility (model exceeded the gateway MCU by 8.2× → rejected; storage-format compression had better ROI) · alert threshold analysis
SecuLetter Inc. — Threat Intelligence Team, Research Institute
2022.04 – 2024.12 (2 yrs 9 mos) · Team of 3
Threat-intelligence AI research and TI service platform development/operations.
- Built the MARS-TI data lake and ETL — extracted visualization data from AWS-crawled raw data and OpenSearch, transforming it into mail/URL/file relationship structures. Built the monitoring UI in Svelte
- Implemented MARS AI for malicious mail classification (70% contribution) — mail preprocessing, BERT-CNN header feature extraction and classification, served via FastAPI + AWS Lambda·ECS·ECR
- Built the training-data crawler (50%) and the internal data inspection platform
ConTI(80%) — Svelte/Tailwind + FastAPI/Lambda/OpenSearch - Improved MARS-TI platform ↔ AI service integration (60%) — connected FastAPI to AWS Lambda over VPC, strengthened API Gateway·ECS·ECR·OpenSearch wiring
Ingest → clean → store → search → API → UI, run end to end by a small team. → Threat Intelligence Data Platform
Education
Hoseo University Graduate School, Computer Engineering (M.S.) · 2020.03 – 2022.02 · GPA 4.09 / 4.5
Thesis: Design and Implementation of a Graph Neural Network-based Behavior Classification Model — wearable 6-axis sensor → LSTM features → GAT node classification, 91.1% test accuracy across 5 behaviors
Hoseo University, Computer Software (B.S.) · 2014.03 – 2020.02
Awards
- Best Graduate Thesis Award — Hoseo University Graduate School, 2022
- Encouragement Award — Korean Institute of Broadcast and Media Engineers, 2019
- 5 funded research projects — NRF ×2, industry-academic foundation ×1, local government ×1, university ×1 (2019–2022)
The full list of papers and talks is on the Research page.