Special · Case study

A neighborhood dentist's AI makeover: from a back-room server rack to AI phones and insurance checks

"A Neighborhood Dental Clinic's AX Transition," posted Oct 13 to the 'AI Local Homelab' channel on Korean forum Arca.live, passed 210,000 views and 140 comments by the morning of Oct 15. A dentist with a biology degree used Claude and Codex coding agents to automate most of his clinic's back office in a few months. This summary follows his own account; all figures are his claims.

1. It started as a complaint to a friend

   The author says he was complaining that AI could draw and code while he was still "bent over someone's mouth doing fillings all day," and a friend told him to try AI himself. He opened ChatGPT that night. The first real step came when he opened his practice: the network wiring quote was so high that he pulled the LAN cables through the ceiling himself. NFC tags on the staff door now log check-ins to an internal portal, which calculates hours and hourly pay automatically.

2. Rule one: patient data never leaves the building

   He decided not to send call recordings or insurance paperwork to cloud AI services. That meant running AI inside the clinic, which in his words meant "buying hardware, and my wallet crying." Attendance logs, calls and consultation records all stay on the servers in the back room.

3. The server rack in the back room

   Twelve NVIDIA DGX Spark-class boxes are cabled together in groups of four to run large models: four on DeepSeek, four on GLM and four for experiments. Two Mac mini M5 Pro machines handle voice, search and document reading, duplicated so one can fail without stopping work. There is also an M4 Mac mini for agents, three virtualization servers and six UPS units. Counting exam rooms and the front desk, the clinic has more than 30 computers. One lesson learned: plugging the high-speed cable in while powered on locks it at a tenth of its speed until you unplug and wait a minute.

4. An AI answers the phone first

   Instead of a carrier's business phone plan, he built his own PBX and an app that lets him make and take calls on the clinic's main number from home. An AI voice picks up first, with 0.1 seconds to first sound. Calls are transcribed live onto staff screens, and afterward the system separates speakers and writes a summary and to-do list. He jokes that he "built a system to exploit myself" and no longer has an end to his workday.

5. Insurance reconciliation paid off most

   The AI read a backlog of 904 documents, 3,389 pages, in 69 minutes with zero failed pages. He tested four document-reading models on real paperwork: on blurry scans, free OCR got 62% of amounts right and the model he chose got 100%. The AI lists each line's amount and denial reason and marks it "confirmed" if it matches clinic records or "needs review" if not. Of 6,928 lines, 5,319 (77%) were confirmed automatically, so staff only check the remaining 23%. One employee's reaction: "Am I getting fired now?"

6. Internal search, messenger and PA system

   Call logs, texts, insurance paperwork, faxes, manuals and even dental textbooks are searchable in one place. Staff can just ask things like "what do we do when an implant claim is denied?" In a test with 28 real staff questions, it found the right answer for 26. He replaced group chats, which buzzed after hours and sent patient talk to outside servers, with a messenger he built himself. When the front desk pages an exam room, an AI voice reads it through the ceiling speaker so gloved staff don't have to look at a screen.

7. The dentist grades the AI's homework

   When the AI summarizes a call and tags it as, say, "reschedule, medium urgency, follow-up needed," he grades it with number keys on a screen he built. Calls where DeepSeek and GLM disagree come first, and a blind tab hides which model wrote each answer so he can pick the better one. The grades become an answer key for comparing models and future fine-tuning on the clinic's own calls.

8. Business reports and bookkeeping

   After each clinic day a summary PDF arrives, with a weekly report on Monday mornings and a diagnostic report on the 1st of each month. For the monthly one, DeepSeek writes the main report and other models add comparison versions; if all three agree, he treats it as a real problem. Numbers are calculated by analysis software, not the AI, which only interprets them and must mark estimates as [estimate]. Bank and card feeds are reconciled against the books, and he can type questions like "why is last month's card revenue 300,000 won off from the ledger?" into a terminal.

9. Paperless consent forms

   Consent forms and new-patient intake moved to tablets. Once staff explain and sign and the patient signs, a PDF is saved to the chart automatically, along with any drawings he makes on the form while explaining a procedure. Many older patients still need paper, though, which he says shows that "the biggest bottleneck in AX is DX, and people."

10. What he cut

    Once AI took over monthly bookkeeping and reports, he canceled his accounting service, though tax filings still go to a human professional. Subscriptions for call recording, OCR and appointment texts were replaced one by one with the back-room servers, and dropping his phone system contract alone saved 4.5 million won. He also cut blog agencies and search ads, bringing marketing spend to zero. Instead he restructured the clinic's website for search engines and AI assistants, and says more patients now arrive saying ChatGPT or Gemini recommended the clinic.

11. How a non-engineer built it

    He says he wrote almost no code himself. He runs Claude and Codex coding agents in separate roles, covering the portal, the phone system, the voice server and infrastructure, and they message each other while they work. Any change to a live setting needs his direct approval, and approval relayed between agents ("the boss said yes") doesn't count. When one agent missed a setting that made name announcements take 6 seconds each, another agent caught it before it went live.

12. Why DGX Spark

    His reasoning was that big models need memory first. An RTX PRO 6000 has 96GB per card and an H200 has 141GB, while one DGX Spark has 128GB and four linked together reach 512GB. Power wiring, heat and noise also mattered. "An H200 is a sports car, a DGX Spark is a one-ton truck," he wrote: slower, but the furniture (a big model) fits. Recipes posted on NVIDIA's forums let a non-engineer run the latest models. He admits it started with wanting one after seeing Jensen Huang hand the box out at a fried-chicken restaurant.

13. His answer to "overinvestment" and "scam" comments

    As the post spread, commenters called it overinvestment or asked what he was trying to sell. He says it is a hobby ("others buy watches, I bought DGXs"), that he has nothing to sell, and that if he releases it he plans to open-source most of it for other dentists. He lists failures too: the voice bot couldn't pronounce weekdays at first, and he had to rebuild the phone system from scratch. "This isn't 'do what I did,'" he wrote. "It's a record that even a non-engineer neighborhood dentist can try this much."

14. Negative@ilpyung98 · X ↗

    Reaction: "1.9 to 3.8 years to pay back, a loss if you count his time"

    X user ilpyung98 estimated the economics using only the hardware and canceled costs mentioned in the post. Assuming about 106 million won in equipment and 56 million won in net annual savings, payback takes 1.9 years, or 3.8 years under conservative assumptions, and any profit disappears once the dentist's own time is priced at a dentist's rate.

15. Mixed아카라이브 댓글 · arca.live ↗

    Reaction: "You can't go AI-native on Dentweb"

    In the comments, a user who said they work in a dental office wrote that Dentweb, the practice software most Korean clinics use, barely opens an API and often blocks database access even on the internal network, so "there's no way to work AI-native like this post with it." Another commenter from the industry called it impressive "but you can't make money with it; it's a hobby."

16. Positive@lucas_flatwhite · X ↗

    Reaction: "This is what AI-native actually looks like"

    On X and in the comments, people wrote that they'd believe it was an IT company rather than a dental clinic and called it "truly AI-native," and several asked him to teach a course for medical professionals.