Jonathan G.

I’m building a room-aware clinical voice system that lets doctors ask questions about the patient in front of them and receive real-time answers from the chart.

The hard part isn’t making an LLM talk. It’s proving the system has the right patient, the right evidence and the right authority before it answers or acts.

CURRENT: VOICE + AGENTIC EMR

For a multi-facility medical clinic in Dallas, I’m leading the AI and EMR architecture behind a purpose-built voice device. It combines room proximity, clinician identity and active encounter data to bind the correct doctor, room and patient before accepting a question.

A doctor can ask, “What was her last A1c?”, “Why was this medication stopped?” or “Did she complete the cardiology follow-up?” and receive a low-latency answer assembled from structured EMR records and longitudinal clinical notes.

The answer is designed to show its evidence, dates and conflicts—and abstain when the chart does not support a reliable conclusion.

The same platform captures the encounter, separates clinician and patient speech, transcribes and translates multilingual conversations, and prepares chart-ready documentation. We’re extending it into governed EMR agents that can identify unfinished follow-ups, prepare referrals and orders, assemble prior-authorization evidence, coordinate scheduling, draft patient instructions and reconcile documentation.

Underneath the interface are FHIR/HL7 integrations, patient and encounter identity resolution, permissioned tool calls and source-level auditability.

The operating model is simple: AI does the searching, assembling and drafting; the doctor reviews and verifies. Read, prepare and act are separate permissions. Consequential actions require approval, an audit trail and verification that the EMR actually changed.

In parallel, I’m building the dialect-rich, human-reviewed Arabic speech data and evaluation pipeline needed to post-train open-weight ASR models. The roadmap includes LoRA adaptation, distillation and quantized serving for lower-cost self-hosting and possible edge deployment.

SELECTED PROOF

• For Plan AI, I built and productionized a real-estate intelligence engine that finds and interprets zoning rules, classifies interior, corner and through lots from geospatial data, calculates setbacks and buildable envelopes, and incorporates FEMA flood risk. Its evaluation harness catches invented ordinances, invalid geometry and unsupported conclusions—allowing cheaper models such as Gemini Flash to be used while still returning verifiably correct results.