PORTFOLIO / CHATBOT & RAGBUILD · OPERATE · IMPROVE

WOON JANG / AI PLATFORM · RAG ENGINEER

I build and operate internal AI systems.

From on-premise chatbot operations to permission-aware document RAG design, I build AI systems for everyday work.

CHATBOT KB-ELEMENTSimulated demo
Public simulated demo: a fictional attachment-summary question, sample answer, history and composer. Screenshot in Korean.
Try the chat UI ↗Explore the architecture ↗

01 / ON-PREMISE CHATBOT

An internal chatbot for attachment-based reasoning and web search

Operating project

Document and image analysis and web research share one conversation interface. A request queue controls inference on limited GPU resources.

What the user sees

Simulation
Enlarged simulated document answer: no file content was read, a sample review table and feedback. Korean UI.
Public demo answer detail · Not real inference
  1. ①InputQuestion · attachment selection
  2. ②AnswerStreaming display · feedback
  3. ③HistoryConversations · reopen

What the server handles

  1. Question & filesReact
  2. Auth & contextFastAPI
  3. Request queuePostgreSQL FIFO
  4. Model inferenceSGLang · GLM-5.3 Flash
  5. Streaming answerSSE → React

API ↔ PostgreSQL: accounts · conversations · files · feedback

A summary of the operating source. The public demo does not execute this server path.

02 / PERMISSION-AWARE RAG

RAG with a centralized internal reference database as the source for answers

Design baseline

Approved organizational documents become searchable knowledge. The design retrieves evidence within user permissions and links answers to their sources.

A. Documents to knowledge

  1. Approved documentsApproval gate
  2. Parse & splitVersion · ACL metadata
  3. Search indexQdrant · PostgreSQL

Indexed knowledge → permission-filtered retrieval below

B. Questions to grounded answers

  1. Question & userTrusted user · groups
  2. Authorized retrievalACL pre-filter
  3. Grounded answerDecline if evidence is insufficient
  4. Source referenceDocument · version · as-of date
Design concept: index approved documents, retrieve permitted evidence and attach sources to answers. Detailed ERD and implementation status are separated in the atlas.
Fictional format example
Question
Which documents are needed for expense reimbursement?
Answer
An application and supporting receipts are required. [1]
Source
[1] Fictional policy · v1.0 · p.3
View RAG architecture ↗

03 / OPERATING EXPERIENCE

Design based on operational experience

Team project

Industrial data pipeline

Processing overview

  1. Ingest & retainWebhook · raw hex
  2. NormalizeConverter · InfluxDB
  3. Batch & exportCelery · Avro/GCS
Evidence & detailed architecture ↗
Solo product development & operations

untamedai.me

Product operating loop

  1. Design & buildNext.js · FastAPI
  2. DeployCloudflare
  3. Operate & improveSessions · cost · safety
Product operations evidence ↗

04 / CONTACT

EmailGitHub ↗Career Summary PDF (Korean)