Seunghun MaOn a call 00:00KO
(00)Call

I work between technology and business.
How can I help?

I lead applied AI projects, and design and evaluate them myself. My work includes a public-agency AI voice agent pilot, speech recognition research, and audits of LLM agents.

FIELDTECHBUSINESS
◦Overview
73.4%
Share of 1,385 valid calls finished by AI
3
Projects as Principal Investigator: Collabo R&D Phase 2, R&D Foundation Support Program, Publishing Content Technology Development
22.54→11.49%
Word error rate, domain speech recognition
12 days
Round-the-clock pilot at a public agency

FIELD

  • Define the problem in the meeting, and settle the AI to use, the success metrics and the plan on the spot
  • Test feasibility on real data during the meeting
  • Build and run an AI voice agent for a public agency, 12 days, 24/7
  • Design that decides first where a person checks

TECH

  • Speech recognition: researcher in Collabo R&D Phase 1, Principal Investigator in Phase 2
  • R&D Foundation Support Program, Principal Investigator: persona feedback loop for an AI web-novel writer
  • Researcher on a manual-automation method for optimizing knowledge-base answers
  • Evaluation method that audits LLM-agent code and LLM grading requirement by requirement
  • Full-stack development and cloud deployment with Python, Node.js, React and FastAPI

BUSINESS

  • Choose and migrate to a cloud that meets public-sector security requirements
  • Find common ground between companies, universities and reviewers in industry-academia projects
  • Publishing Content Technology Development under way; as Principal Investigator, handle the performance test
  • Decide when to hold a release or accept a lower level of performance
(01)

Listen

I start by listening to what the field and the customer need.

Define the problem in the meeting, and settle the AI to use, the success metrics and the plan on the spot
When needed, test feasibility on real data right there
In a voice conversation, a slow reply makes callers think something is broken, so cutting delay matters most
As Principal Investigator on Collabo R&D Phase 2, research real-time speech recognition that writes text before the speaker finishes
RECORD 01 · Live transcription
Audio VAD Buffer STT Commit Text Audio VAD Buffer STT Commit Text
(02)

Decide

First, I sort out what will work and what won't.

Leave a method out when experiments show no effect
The simpler method is the answer
Audit AI-written code and grading requirement by requirement; cover what technology cannot with human check points and agreements
RECORD 02 · Decision log
  1. 01Is keyword biasing still needed after fine-tuning?Left outChecked after Collabo R&D Phase 1. On one domain, 3,000 sentences, the +0.3 pt gain was not significant and precision dropped.
  2. 02A Python process started for every callSimpler optionA process pool meant more to manage, so I ported it to Node.js instead. Memory use fell to a quarter.
  3. 03Public-sector security requirements and the cloudMovedTwo major global cloud providers fell short of the required level at the time, so I moved to a CSAP Level 2 certified cloud.
  4. 04Can DB-write code from an LLM agent be trusted?Audit by requirement22 adapters audited against 9 requirements. The spec covered 1 of 9, and the checks observed 0 of 9.
  5. 05Does LLM grading add anything beyond surface features?Tested separatelyTested apart from surface features such as character count and grade level. No statistically detectable added effect from GPT-4o.
  6. 06Review steps where AI errors remainPhased rolloutAgreed to widen automation in stages rather than hand everything over at once.
(03)

Answer

I build it to hold up in real use.

Ran an AI voice agent for a public agency for 12 days, 24/7
Combined telephony, speech recognition, a language model and speech synthesis into one system
Moved to a cloud certified under Korea's CSAP, Level 2, to meet public-sector security requirements
RECORD 03 · Pilot calls, 1,385
0.0%of valid calls finished by AI · target 70%
Finished by AI 1,017Handed to a person 368
CSAP Level 2 certified cloud Phone line PBX STT LLM TTS CSAP Level 2 certified cloud Phone line PBX STT LLM TTS
(04)

Connect

When a person is needed, I connect you.

Find common ground between the companies, universities and reviewers in industry-academia projects
When a request is hard, explain why and offer an alternative that works
Keep a record of what blocked us; a carrier registration failure was solved that way
Projects as Principal Investigator: Collabo R&D Phase 2, R&D Foundation Support Program, Publishing Content Technology Development
Speak to a person 0 Seunghun Ma
RECORD 04 · The handoff
00:03AIHello. How can I help?
00:07CallerI'd like to talk to Seunghun directly.
00:11AIUnderstood. Connecting you to Seunghun now.
00:12●Seunghun connected
(05)
00:00

The call has ended.