AI SPACE · SYSTEM DESIGN & DELIVERY

AI writes the code.
30 years of experience does the design.

Thirty years designing network, VoIP and IMS systems; the last eight on AI/LLM and voice. These days I write the design document first, hand implementation to AI agents, then verify the result myself before it ships. I do not design and leave.

Telecom · VoIP · IMS, 30 yrsAI/LLM · voice, 8 yrs100+ projectsPatent filed, 5G voice agent

01 · Why design, now

Plenty of people can code well.
Few can design well.

With AI coding tools, writing code fast is no longer the problem. The problem is on either side of it: deciding what to build and in what structure, and checking that what got built runs the way the design says. That is where vibe-coded systems have broken every time I have seen one.

Code piled up without structure

Add features one prompt at a time and in two or three weeks nobody can touch the code. Unless table design, module boundaries and state transitions are written down first, AI-written code does not stay maintainable.

Shipping without verification

Running and running correctly are different things. An AI writes tests that its own code passes. Edge values, failure recovery and permission flows need a person to decide the test items and a person to read the results.

Turning requirements into structure

What a client hands over is a feature list, not a system. Turning it into entities, flows and operating procedures is the design work, and that is where experience shows.

02 · Order of work

The design document comes first. Agents do the implementation.

I kept the way I have designed systems for 30 years and changed only who writes the code. Once structure and data model are fixed in a design document, agents with separate roles implement from it. I hold the verification gate.

Design–implement–verify process The architect fixes requirements and structure, a master agent distributes work to implementation, review and test agents, and the architect verifies the result before delivery. AI AGENTS SPEC OUTPUT REVISIONS Architect needs · structure · data Master agent dispatch · coordinate Build code gen Review spec conformance Test automated Verify · Deliver test report · spec diff Human (architect) gate AI agent
01 Design · human

Fix the structure first

Needs become entities, states and flows.

  • Data model · ERD
  • Permissions · operating procedures
  • Failure and recovery scenarios
02 Build · agents

Implement by role

A master agent dispatches work against the spec.

  • Build agent: code generation
  • Review agent: spec conformance
  • Test agent: automated tests
03 Verify · human

Check against the design

Not whether it runs, but whether it runs as designed.

  • Test item design
  • Edge cases · load tests
  • Revision loop until it passes
04 Deliver

Hand over with documents

Never code alone.

  • Design doc · operations guide
  • Test result report
  • Defined maintenance scope

How the agents are split is design work too. Role split, where a human checks, and what goes to automated tests are decided again for each project. The same setup is not reused everywhere.

03 · Case studies

Three design cases

Client names are left out. Each case covers what the structural problem was and how it was solved.

IoT door control service

MQTTAWS IoT CoreiOS · AndroidPostgreSQL
Problem
When a device changes owner through a move or resale, the previous owner's family can still open the door. No recovery path when a user replaces their phone.
Design
A device-centric permission mapping (1 admin · N users) separates roles, with a lifecycle that returns a device to an "inactive" state. On ownership transfer, all subordinate user links are cleared at that moment. A planned phone swap self-migrates through a pre-announced handover window; only loss or breakage needs manual operator recovery.
Result
Recovery operations became simple without SMS verification, and operator intervention was limited to loss and breakage.

Hospital medical chatbot guardrail

LLMRAGClassifierOn-premises
Problem
A patient-facing chatbot could emit unsupported medical answers. Prompt instructions alone could not control it.
Design
A two-stage structure with lightweight classifier gates before and after the LLM. Query classification decides the answerable scope first, and an answer passes only when the hospital knowledge base supports it. Patient data stays on the hospital's own GPUs.
Result
Guardrail policy survives model changes. The hospital's review items were turned into the classifier dataset.

5G private-network AI voice agent

IMS · SIP5G NPNSTT · TTSPatent filed
Problem
Attaching an AI voice agent to a carrier IMS core requires giving the agent a subscriber identity, and a fixed identity tangles security and billing responsibility.
Design
The agent receives a temporary IMS public identity, and a "trusted delegate" model limits its call-handling authority to what the real subscriber delegates. The identity is revoked when the session ends.
Result
AI agent call handling without modifying the existing IMS core. Patent filed on the architecture.

04 · Services

From design through testing, I do it myself

Design document, code and test report are delivered as one set. Maintenance scope after handover is agreed separately at the quote stage.

AI voice agents · SIP/VoIP gateways

SIP · RTP · STT/TTS · LLM

AI chatbots · RAG knowledge bases

LLM · RAG · guardrails · on-prem

IoT · device control services

MQTT · BLE · AWS IoT · ESP32

Web · mobile · backend systems

KMP · FastAPI · PostgreSQL · Elasticsearch

Workflow automation · agent pipelines

multi-agent · n8n · email integration

AI infrastructure design

GPU servers · vLLM · Kubernetes

Documents over meetings

Requirements note, design document, test report: each step is confirmed in writing before the next. Meetings only when they are actually needed.

Scope gets cut first

Phase one is a narrow MVP; the rest goes into upgrade items. I do not quote everything up front.

Responsibility boundaries go in the quote

Development vs. operations responsibility, the security level applied and the maintenance scope are written into the quote, so there is no argument later about who owned what.

Independent contractor

Working with clients in Korea and abroad. Remote, asynchronous engagement by default; English in writing.

05 · Contact

Requirements do not need to be sorted yet

Email me what you want to build, in whatever shape it is in. I will propose scope and structure first; the quote comes after.

Jongchul (Tony) LimAISPACE LAB · SYSTEM ARCHITECT
EMAILjclim8009@gmail.com
PHONE+82 10-9141-8009