← INDEX/WORK/DOPPELDESK

№ 07 — VOICE AI PLATFORM

Doppeldesk

The PTC AI Operator, made into a product. A business connects its recorded calls; Doppeldesk clones its best operator, takes the after-hours line first, and reports booking rate against its own human baseline.

TYPEVenture · SaaS
YEAR2026
ROLEFounder · Architect · Engineer
STATUSLive · First tenant onboarded (Professional Travel Center) · Founding 10 open
SCOPECall mining · Voice + chat agents · Drift & contract gates · Multi-tenant SaaS
LIVEdoppeldesk.com ↗

Why this exists

The PTC AI Operator proved an AI could answer a real marketing line and book real appointments. The question it left behind was whether the hard parts generalize: mining a company’s own calls into the script, gating every change so what was tested is what answers the phone, and measuring against the humans it relieves. Generic AI receptionists start from a blank form; voice platforms hand over a toolkit. Doppeldesk starts from the customer’s own recordings and ends with a number against their own operators. Five stages, one pipeline: Listen → Learn → Launch → Prove → Renew.

The first tenant is the company the prototype was built inside. Professional Travel Center was seeded in October 2026 with 962 real inbound calls from its phone system over fourteen days and 263 AI-agent calls, synced every twenty minutes by a database cron. Its team works the same Call Activity screen the product ships: reports by hour and day, filters by operator and number, recordings, tags, one-click transcription. The lead segment is sales-driven call rooms, where the call is the sale: direct-response, membership and seminar lines. Pricing is public: Launch at $249 a month, Clone at $399, Campaign line at $699, with the first ten businesses getting setup waived and the price locked for a year.

The proof is deliberately honest. The pilot line takes only rollover and after-hours calls, so the headline is recovered calls, not a contest with the operators: in five weeks the AI desk answered 246 calls no operator was free to take and set 67 appointments from them, which became 24 shows and 4 sales. On mail a week old or less it set 47–52% of verified calls; the operators book 42% of every call that comes in. The board separates the operator from the lead source, and the product never claims to beat the humans before a true cutover produces like-for-like data.

What I engineered

LISTEN

Connectors, not uploads

RingCentral / Office@Hand with JWT refresh, detailed call legs and recording fetch; CallRail; iovox CSV; MP3 upload. Every inbound call lands in a tenant phone log; recordings stay in the tenant’s bucket; ElevenLabs Scribe v2 transcribes diarized on demand. Nothing crosses tenants.

LEARN

The miner

Claude Haiku labels outcomes and extracts signals from each call (questions, objections, the winning line); Claude Opus synthesizes the playbook; the human booking baseline is measured from the log before the AI takes a call. Industry packs hold aggregate patterns only, never transcripts. The first pack is direct-response events, learned from the pilot.

LAUNCH

One answering rule

Voice and web chat from one agent definition on ElevenLabs Agents, Claude reasoning inside. Rollout is supervised by default: after-hours → overflow → full line, flipped by a phone-system rule. Recording-consent line chosen by caller area code, strictest form when unknown; AI disclosure on every greeting. The onboarding wizard goes site → draft → questions → build → “Talk to your desk”.

GATES

Lint, contract, regression, drift

A renderer turns pack + playbook into the system prompt, tools.json and expected agent config. The contract gate fails the build if any reason the API can return lacks a prompt branch. Regression runs simulated conversations seeded from mined objections. Nightly drift diffs the live agent against the repo; publish() refuses a version that failed.

GUARD

The live-agent guard

A database cron probes every live agent each minute through pg_net, so the API can stay asleep. On a block it alerts, then restores the last version proven by a real conversation, once per incident. 1,440 of 1,440 checks passed in its first day. A version is “proven” only when a caller actually talked to it.

PROVE

A receipt-true board

A call counts as booked only when a booking row carries its conversation id. The owner sees booking rate against the human baseline, a stage timeline (Listen / Learn / Launch / Prove), playbook evidence counts and version history with gate results. The Call Activity screen gives call-tracking parity: tiles, by-hour charts, filters, export, recording playback, tags, notes.

SCAN

The call scan

Quick and deep scans over a business’s phone log: a switchboard jack grid lights in the real answered / missed proportions while findings drop in with their evidence counts. Paced playback of real scan events only, never a canned animation. Also ships as a CLI for an offline first look.

TENANCY

One database, RLS everywhere

tenant_id and row-level security on every table. Dashboard sessions are read-only; every write goes through an admin route under the service role, so no one can flip a gate result. Capacity lives in SQL functions, never in the prompt. Agent-facing routes never return 4xx/5xx to the agent’s own mistake: always 200 with a reason the prompt knows how to handle. No timers in the process; schedules live in pg_cron.

VERIFIED

38 live checks after every backend change

One script exercises the live demo line end to end: agent tool dry-runs, consent by area code, signed post-call ingest with audio and outcome labels, every job, cron picking up a queued scan on its own, admin routes with 401/403/RLS, billing, then deletes everything it created. 38 of 38 pass.

In the wild

Doppeldesk platform pipeline — Listen, Learn, Launch, Prove, Renew, with the five gates every change passes and the supervised after-hours-first rollout
FIVE STAGES, ONE PIPELINE — LISTEN → LEARN → LAUNCH → PROVE → RENEW, FIVE GATES, AFTER-HOURS FIRST
Pilot results on the first tenant — 246 rollover and after-hours calls to 67 appointments, 24 shows and 4 sales, set rate by mail age, against the operators' 42% bar
THE PROVE STAGE — 246 RECOVERED CALLS → 67 APPOINTMENTS, SET RATE BY MAIL AGE, THE OWNER'S OWN BAR
246 → 67RECOVERED CALLS → APPOINTMENTS, 5 WEEKS
962REAL CALLS IN THE FIRST TENANT
1,440/1,440GUARD CHECKS, FIRST 24 HOURS
38/38LIVE BACKEND CHECKS PASSING
ElevenLabs AgentsClaude (Anthropic) · Sonnet on calls, Haiku + Opus in the minerExpress · TypeScriptNext.js 15Supabase Postgres · RLS · pg_cron · pg_net · VaultRingCentral · CallRail · TwilioElevenLabs ScribeRender · NetlifyResendRemotionClaude Code

The step from forward-deployed build to product: the parts that made one AI operator trustworthy, mined scripts, gates and a baseline, became a platform any phone-driven business can onboard, with the original customer as tenant one and its numbers reported the way an auditor would want them.