I'm working through the KANT Private LLM Engineer program, step by step from machine-learning fundamentals to RAG, AI agents, LLMOps and model serving. I learn by building the pieces LLMs need in real work โ document search, information extraction and API servers. ๋ถ๋์ฐํ ์ ๊ณต ๊ฒฝํ์ ์ด๋ ค, ๋๋ฉ์ธ ์ง์์ด ํ์ํ ๋ฌธ์ ์ LLM์ ์ฐ๊ฒฐํด ๋ณด๋ ๋ฐ ๊ด์ฌ์ด ๋ง์ต๋๋ค.
์ง๊ธ์ LLM์ ์ค์ ์๋น์ค๋ก ์ฐ๊ฒฐํ๋ ๊ตฌ์กฐ์ ์์ ์ ์ธ ์ด์ ๋ฐฉ๋ฒ์ ์ง์คํ๊ณ ์์ต๋๋ค.
- LLM application structure โ LangChain Prompt โ Model โ Parser, structured output and validation
- Backends for LLMs โ designing document, query and Q&A APIs with FastAPI + Pydantic
- Reliable external calls โ timeouts, bounded retries, backoff + jitter, fallbacks
- Reproducible experiments โ baselines, train / validation / test splits, leakage checks, experiment logs
- Deployment โ AWS DevOps course (๋ฐฐํฌ๊น์ง ํด๋ด๋ ๊ฐ๋ฐ์) in parallel
| Area | Stack |
|---|---|
| Language | |
| Data ยท ML | |
| Deep Learning | |
| LLM ยท Backend | |
| Tools |
์ฒญ์ฝํ ์ ์ฃผ์๋ชจ์ง๊ณต๊ณ ๋ฌธ PDF๋ฅผ ์ฝ๊ณ ๋ถ์๊ฐยท์ผ์ ยท๋์ถยทํน๋ณ๊ณต๊ธ ์กฐ๊ฑด์ ๋ตํ๋ ๋ก์ปฌ LLM ์ฑ๋ด์ ๋๋ค.
- Collected 30 notices and built a 20-item evaluation answer set (schema design โ review โ audit)
- Numbers are calculated in code, not by the LLM, and every answer cites the original sentence from the notice
- Local serving with FastAPI + Ollama in Docker, designed so vLLM can be swapped in
- Comparing 20 ways to read PDF tables (4 rule-based ยท 16 LLM-based)
- ๊ณต๊ณ ๋ฌธ๋ง๋ค ํ ํ์์ด ๋ฌ๋ผ์, ์ด๋ค ๋ฐฉ๋ฒ์ด ๊ฐ์ฅ ์ ํํ์ง ์ง์ ๋น๊ตํ๊ณ ์์ต๋๋ค.
Keyword baseline๊ณผ TF-IDF ๊ฒ์์ ๊ตฌํํ๊ณ Precision@3, MRR๋ก ์ฑ๋ฅ์ ๋น๊ตํ ํ๋ก์ ํธ์ ๋๋ค.
- Text preprocessing and TF-IDF vectorization
- Cosine similarity implemented with NumPy
- Search evaluation set with Precision@3 and MRR
- Failure-case analysis and title-weighting experiments
๋งค๋ฌผ ๊ด๊ณ ๊ฐ ๊ตญํ ๊ตํต๋ถ ํ์ยท๊ด๊ณ ๊ธฐ์ค์ ์ด๊ธ๋๋์ง ํ์ ํ๋ ์ด์์คํดํธ๋ฅผ ๊ฐ์ ํ๊ณ , ๋ง๋ ๋ก์ปฌ ๋ชจ๋ธ์ ์คํ์ผ๋ก ๊ณจ๋์ต๋๋ค.
- Compared
gemma3:4bandqwen3:4b-instructat the same size (4B) and quantization (Q4_K_M) - 10 questions (normal ยท edge ยท out-of-scope), 40 runs in total, measuring quality and speed (RTX 5060 Laptop, 8 GB VRAM)
- Answer accuracy: gemma3 57.0% vs qwen3 80.8% โ selected
qwen3:4b-instruct - Ran the same questions against a cloud API to check whether local operation is viable
KANT ๊ณผ์ ์์ ๋ฐฐ์ฐ๊ณ ์ง์ ๊ฒ์ฆํ ๋ด์ฉ์ ๊ธฐ๋กํฉ๋๋ค. Exercises are marked done only when I have run them and can explain them.
| Track | Coverage |
|---|---|
| Machine Learning | Model evaluation, ensembles, biasโvariance, imbalance & cross-validation, CV tuning |
| Deep Learning | PyTorch, MLP ยท CNN ยท RNN, Attention, tokenization, fine-tuning, PEFT |
| LLM in Practice | Logits ยท loss ยท optimizers, reliable LLM API calls |
| Data Engineering | API collection, cleaning & RAG document sets, FastAPI ยท Pydantic, async calls |
| LangChain | App structure, prompt templates, output parsers, LCEL, retrievers |
flowchart LR
A[Machine Learning] --> B[Deep Learning] --> C[Transformer & LLM] --> D[LLM Application]
D --> E[RAG] --> F[AI Agent] --> G[LLMOps] --> H[Private LLM Serving]
classDef done fill:#dcfce7,stroke:#16a34a,color:#14532d
classDef now fill:#fef3c7,stroke:#d97706,color:#78350f
classDef next fill:#f3f4f6,stroke:#9ca3af,color:#374151
class A,B,C done
class D now
class E,F,G,H next
- B.A. in Real Estate, Konkuk University, Seoul (๊ฑด๊ตญ๋ํ๊ต ์์ธ์บ ํผ์ค ๋ถ๋์ฐํ๊ณผ)
- ๋ถ๋์ฐ ๋ฒ๊ทยท๊ณ ์๋ฅผ ์ฝ๊ณ ํ๋จ ๊ธฐ์ค์ผ๋ก ์ ๋ฆฌํ๋ ๋๋ฉ์ธ ์ดํด โ ํ์๋งค๋ฌผ ํ์ง ์ด์์คํดํธ์์ ๊ตญํ ๋ถ ํ์ยท๊ด๊ณ ๊ธฐ์ค์ ํ๊ฐ ๋ฌธํญ์ผ๋ก ์ค๊ณ
- KANT Private LLM Engineer program
- ๋ชจ๋ฅด๋ ๋ถ๋ถ์ ์ถ์ธกํ์ง ์๊ณ ์ง์ ํ์ธํฉ๋๋ค.
- ๋์ ์ซ์ ํ๋๋ณด๋ค ๋ฌธ์ ๋น์ฉ์ ๋ง๋ ํ๊ฐ ๊ธฐ์ค์ ๋จผ์ ์ ์ํฉ๋๋ค.
- ๊ฒฐ๊ณผ๋ฟ ์๋๋ผ ๊ฐ์ , ์คํ ๊ณผ์ ๊ณผ ์คํจ ์ฌ๋ก๋ฅผ ํจ๊ป ๊ธฐ๋กํฉ๋๋ค.
- ์ค๋ฅ๋ฅผ ์ฌํํ๊ณ ์ฌ์ฉ์ ํผ๋๋ฐฑ์ ๊ฐ์ ์ผ๋ก ์ฐ๊ฒฐํ๋ ์ด์์ ์งํฅํฉ๋๋ค.


