A voice system that only confirms what really happened.
DentSignal connects calls, speech, booking, records, and handoff without hiding failed work.
My work
System design, backend, voice orchestration, reliability, operator experience
Core stack
Python, FastAPI, PostgreSQL, React, voice and telephony integrations
Evidence
Real interface; synthetic fixtures; historical bounded regression runs
Status
Independent engineering work—not a client deployment claim
One call, without the architecture wallpaper
The useful path is the one the bytes actually take.
Patient traffic uses Azure ACS. The Telnyx lane was a separate no-PHI sales/self-call experiment. They are intentionally not drawn as one magical telephone cloud.
Separate guarded lane: Telnyx → no-PHI sales/self-call smoke tests → same backend safety gates. Not the patient-call path.
Why these repairs are here
Three bugs exposed three trust boundaries.
These examples show how DentSignal protects call quality, honest booking numbers, and server memory. Each one pairs the broken behavior with the code that fixed it.
Repair 01 · Call audio
Do not send audio before the speech service is ready.
What was broken
Audio was downsampled too early.
The bridge forced higher-rate audio to 8 kHz and could send before the speech session was ready.
What changed
Preserve rate and wait.
Keep the source rate, buffer briefly, then flush only after settings are accepted.
What it fixed
No premature audio send.
An open socket is no longer treated as a ready speech session.
Settings acceptance now gates buffered-audio flush. Mocked provider paths and synthetic audio only; no claim about real PSTN jitter, room noise, listening quality, or production deployment.
Repair 02 · Booking numbers
Count booking success against real booking intent.
What was broken
The denominator changed the story.
Booking rate divided by every call and fixed multipliers looked like recovered revenue.
What changed
Definitions moved into one source of truth.
Use appointment intent as the denominator and label configured values as modeled.
What it fixed
The UI can no longer overstate what storage proves.
Labels, fields, numerator, and denominator now describe the same event.
completion = booked / appointment_intent
modeled_value = sum(call.value for call in booked_calls)
What the bounded retest established
Booking-rate definitions now stay tied to booking intent and basis labels. The current data model still cannot attribute a later booking to an original missed call.
Repair 03 · CSV uploads
Reject oversized files before reading all of them.
What was broken
The size check happened after the cost.
The route loaded the whole file before checking its 10 MiB limit.
What changed
Read only enough to decide.
Read the limit plus one byte before parsing.
What it fixed
Application memory now respects the application policy.
Oversized files stop with HTTP 413 before an unbounded read.
Broken
content = await file.read()
if len(content) > MAX_CSV_SIZE:
reject()
The application reads only enough bytes to decide whether the file is over the limit. This supplements reverse-proxy limits; it does not replace edge enforcement.
Evidence boundary
Engineering proof, not customer proof.
Independent engineering work with synthetic evidence. No claim of clinic outcomes, production uptime, or certification.
