Wireframe — Estimate Recording (Phase 2)

The recording feature lets a tech hit "Record" when they walk into a customer's house, capture the full estimate visit, and have the system auto-transcribe and extract the scope of work. The intake form is pre-filled when they hit "End", so they can review on their phone, tweak if needed, and generate the estimate before leaving the customer's driveway. Phase 2 — ships after the Phase 1 web app.

{{ name }} — Estimate Recording

Phase 2 · Future D Dan ▾

During Visit · active recording

Smith — Bath Visit Recording in progress · session 1

REC

🔔 Reminder: Tech announces "I'm recording this for accuracy." One-party consent in CO — legally fine.

12:34

Capturing audio Mic: iPhone built-in · 47 MB so far
Auto-saves every 30s · safe if app backgrounds

⏸ Pause

⏹ End Recording & Process

After "End" · form auto-filled

Review & Generate
From recording · 12:34 captured

Pre-filled

Job Type ★ AI

Client Name ★ AI

Phone ★ AI

Address ★ AI

Scope of Work ★ AI

Special Conditions ★ AI

Finish Level

Default · adjust if needed

Why This Matters

Tech writes nothing during the visit — full attention on the customer.
Estimate ready before they leave the driveway — sign on the spot.
Audio is the source of truth — disputes have a real record.
Transcript feeds the training pool over time.

What Gets Captured

Audio file → Convex Storage · auto-deletes after 90 days
Transcript → kept forever · attached to estimate
Extracted fields → name, phone, address, scope, conditions
Photos (if tech also taps photo button during/after) → still uploaded the existing way

Cost Per Visit

Whisper transcription: ~$0.006/min
45-min visit: ~$0.27
GPT scope extraction: ~$0.02
Total per recorded visit: ~$0.30

PHASE AIn the Field · Mobile App (Capacitor)

STEP 1

Tech arrives

Opens app, navigates to "New Recording"

→

STEP 2

Tap Record

Native mic permission already granted (one-time setup)

→

STEP 3

Disclosure

Tech announces "I'm recording for accuracy"

→

STEP 4

Visit conversation

15–60 min · audio chunks auto-save every 30s

→

STEP 5

Tap End

Tech confirms · file finalized + queued for upload

PHASE BProcessing · Cloud (n8n + OpenAI)

STEP 6

Upload audio

→ Convex Storage · 5–15s for typical 45min visit

→

STEP 7

Whisper transcribes

~30–60s · returns full text + timestamps

→

STEP 8

GPT extracts scope

~5s · maps transcript to intake form fields

PHASE CReview & Generate · Mobile App

STEP 9

Form auto-fills

All fields highlighted ★ AI · tech sees results in ~1 min total

→

STEP 10

Tech reviews

Edits any wrong fields · adds finish level if model didn't catch it

→

STEP 11

Generate estimate

Hits the existing Generate button · same pipeline as manual intake

→

STEP 12

Show customer

Doc Preview tab · sign on the spot

Web vs Mobile App Wrapper · why both

Web (Browser) Phase 2a · partial

The recording UI works in any modern browser using MediaRecorder + getUserMedia. Fine for desktop or short recordings, but with mobile gotchas.

✓ Zero install — works in Chrome / Safari / Firefox
✓ Identical UI to the production web app
✓ Fine for desktop testing or short visits
✕ iOS Safari pauses recording when screen locks or app backgrounds
✕ Tab switching kills the stream
✕ Battery drain on long recordings

Mobile App via Capacitor Phase 2b · ★ recommended for field use

Same React codebase wrapped in Capacitor → ships to App Store / Play Store with a native audio recorder plugin. The recording survives backgrounding, screen lock, and app switching.

✓ Same React code as the web app — no second codebase
✓ Native audio plugin survives backgrounding + screen lock
✓ Camera capture for photos works natively
✓ Push notifications when transcription completes
✓ Installs from App Store / Play Store like a real app
✕ ~3 days to set up + deploy to stores · then iterate the same as web

Decision: Phase 2 ships as Capacitor-wrapped mobile app + matching web fallback. Both flow into the same Phase B/C cloud processing — only Phase A (the recording itself) is mobile-app-specific.

Technical decisions

RETENTION Audio file auto-deletes after 90 days. Transcript stays forever attached to the estimate (audit trail + future training data).
CONSENT Tech announces recording at the start ("recording for accuracy"). Colorado is one-party consent for in-person conversations, so legally compliant. Disclosure shown as a reminder banner in the recording UI.
RESILIENCE Audio chunks save to local storage every 30s during recording. If the app crashes mid-visit, the partial recording is recoverable — the tech doesn't lose 30 minutes of conversation.
TRANSCRIPTION OpenAI Whisper API — best accuracy, cheap (~$0.006/min). Alternatives: Deepgram (faster), AssemblyAI (more diarization). Default: Whisper.
EXTRACTION GPT prompted with the transcript + the intake-form schema. Returns JSON matching the form fields. ★ AI markers on each pre-filled field signal "review me."
SCHEMA New Convex tables: estimateRecordings (audio file ref + transcript + extracted JSON), linked to estimateSessions. Audio retention is a scheduled job on Convex.
AUTH Capacitor app uses Clerk's mobile SDK. Same Organizations primitive as the web app — multi-tenant from day one.

What's not in scope for Phase 2

Real-time streaming transcription. The first version is "stop recording → process all at once." Streaming during the visit is a Phase 2.5 enhancement.
Speaker diarization ("who said what"). The transcript is one big blob initially. If accuracy suffers, we layer in diarization later.
External Bluetooth mic auto-pairing. Native plugins handle this with default OS-level mic selection. Custom mic switcher is future work.
Customer-facing recording playback. Customers don't see the audio; it's an internal tool.

Wireframe — Estimate Recording (Phase 2 · Future Feature)

Web (Browser) Phase 2a · partial

Mobile App via Capacitor Phase 2b · ★ recommended for field use

Technical decisions

What's not in scope for Phase 2