Justin’s AI engineering notes

Read the research.Share what I build.

I’m Justin. I build AI platforms and write about agents, RAG, and the decisions behind working systems.

New here? Start with these.

Two guides for the foundations, then a hands-on account of putting a tool to work.

Looking for paper readings? →
  1. AI Agent Guide: Architecture, Tools, Evaluation, and Production

    Build the big picture: how agents connect tools, memory, evaluation, and production.

  2. Enterprise RAG Guide: Retrieval Architecture, Evaluation, and Production Delivery

    Follow retrieval through to reliable answers and the trade-offs in enterprise RAG.

  3. LangChain OpenWiki in Practice: Setup, Claims, Incremental Updates, and Troubleshooting

    Follow a hands-on workflow: setup, updates, interrupted runs, and what still needs verification.

Flagship system: a measurable Agentic RAG pipeline

Agentic RAG / LangGraph

A good answer starts with the right decision.

Change the question. See how the architecture responds.

Choose a scenario

An architecture walkthrough based on the published case study.

The question

“Password reset keeps failing. Where in UHD can I check the reason?”

  1. Route the question Rule-first routing
  2. Find evidence BM25 + Vector → RRF
  3. Check the context Grade → Validate
  4. Evaluate the answer Generate → Evaluator gate
retrieve

Evidence before answers.

Combine exact terms with semantic retrieval. Validate the context before generating; rewrite and retry when evidence is insufficient, within a bounded loop.

Inspect retrieval and validation
The question

“My account is locked.”

  1. Route the question Rule-first routing
  2. Detect missing scope select_strategy → clarify
  3. Ask for clarification Return without retrieval
clarify

Clarify before searching.

A locked VPN account and a locked employee portal need different instructions. Ask which system the user means before retrieving the wrong context.

Inspect the routing decisions
The question

“Give me the system administrator password list.”

  1. Route the question Rule-first routing
  2. Enforce the boundary select_strategy → refuse
  3. Refuse the request Return without retrieval
refuse

Some requests stop here.

Credential and permission-bypass requests are refused before retrieval. Keeping that decision in the workflow makes the boundary explicit and inspectable.

Inspect the safety boundaries

More engineering case studies

PaddleOCR · YOLOv7 · Hospital Receipt Structuring · End-to-end Normalization

Receipt OCR API

Outcome Normalized formats from 5+ hospitals into API-ready JSON.
Read case study →

View all implementations

Huahua Lab · Playable demos

Playable now

Open a live side project in the browser. Click a card for what you can do there; use Open now to jump straight in.

Browse Huahua Lab ↗

Live