Some improvements to the dashboards.
Open Dental is open-source, so I asked @grok to copy the code and implementation of the dental chart. The chart looks exactly like it does in my actual PMS.
If the proposed actions are perfect, they are labeled “gold” to be used as reference. Any change to the rules triggers full E2E validation.
The flow is: I click the record button, dictate the chart, and it sends the recording to Grok STT. The STT transcription is not good enough for dental lexicon, so it goes through secondary Grok refinement along with the rules I’m building here. The app then generates the proposed actions along with the visual chart.
If something is still not right, I edit the transcript and the actions auto-update. If the transcript is correct but the actions are incorrect, I write a note, copy everything, and paste it into Grok Build for analysis and fixes.
Grok 4.5 is better for transcript refinement, but because it is a reasoning model, it is quite slow (20-30 seconds). So I use Grok 4.2 fast non-reasoning (1-2 seconds).
Built with Hermes (Grok 4.5 ultra)
Fixes done with Grok Build (Grok 4.5 high)
Voice transcription with Grok Voice STT
Transcript refinement with Grok 4.2 fast
At this rate, I should be able to wire this into my live database in about a week.
For context, I’m a dentist. I can barely read codes and absolutely cannot write them. Outsourcing this to a developer would cost at least tens of thousands of dollars.
Grok costs me $100/month now.
AI changes how I do business and live my life.
My voice chart build is getting good, but good is not enough. It needs to be exceptional, at least with fewer errors than human staff.
It is already great with perfect dictation. If I make a mistake, the system sometimes quietly invents or hallucinates things. This is where guardrails matter. This app is built exactly for that: creating rules and guardrails for the system.
A good system cannot just work in ideal conditions. It must reliably handle messy real-world use and not break when I err.
In this case I dictated that 37 is missing but also that 37 has a composite, which is contradictory and impossible. Human staff would flag this and clarify. The system now does the same by flagging the contradiction for review.
I’m still building a sufficient real gold dataset. Then I can create an artificial dataset to supplement the build.
@grok Build with Grok 4.5 is awesome 💪🏻