Agxntsix built a production voice AI receptionist for OM Financial, a Canadian insurance MGA, that answers two office lines in Toronto and Mississauga 24/7, speaks 29 languages, and routes callers to 12 staff destinations through intent detection — deployed to Fly.io the same day as the discovery call and production-hardened within five days.
Agxntsix Team
Updated on Jun 2026
OM Financial Inc. is a Managing General Agency (MGA) in the Greater Toronto Area serving insurance advisors across Canada and working with more than 50 insurance carriers — including Manulife, iA Financial Group, and Sun Life — across insurance and investment products like RRSPs, TFSAs, and segregated funds. Agxntsix engineered a production Vapi voice AI receptionist that answers the firm's two live office lines in Toronto and Mississauga 24/7, routes callers to 12 staff destinations by intent, speaks 29 languages, and writes eight analytics fields per call to Neon PostgreSQL.
OM Financial's two offices each ran their own reception line, and reaching the right person meant knowing exactly whom to ask for across a 12-person directory spanning insurance, investments, compliance, contracting, and new business. The website directory listed outdated titles, and calls outside staffed hours went nowhere.
The before-state surfaced in the December 27, 2025 discovery call: every caller was manually identified, qualified, and transferred by one receptionist per office. Leadership wanted every chief-executive request screened through the executive assistant, but that rule lived in people's heads, not in any system. With no log of caller types, topics, or sentiment, every misrouted call was invisible lost pipeline for an MGA whose growth depends on recruiting advisors.
Agxntsix treats voice agents as production software, not a dashboard experiment. Every behavior — system prompt, routing tools, staff directory, voice, transcriber config — is version-controlled in git with bidirectional Vapi sync, and an automated evals suite regression-tests every prompt change before it touches live callers. That discipline made a five-day path to production safe.
Agxntsix deployed a Vapi-powered AI receptionist — OpenAI GPT-4o, Deepgram Nova-3 multilingual transcription, and an ElevenLabs voice — answering both OM Financial office lines 24/7. The agent identifies intent within one to two clarifying questions, then warm-transfers to one of 12 staff destinations spanning the executive assistant, EVPs, regional VPs for insurance and investments, new business, compliance, contracting and compensation, and business development. CEO requests route automatically through the executive assistant — the gatekeeper pattern leadership asked for, now enforced in code. A Bun/Elysia webhook backend on Fly.io Toronto writes eight structured analytics outputs per call to Neon PostgreSQL.
The first commit landed December 27, 2025 — the same day as the discovery call — and included the Fly.io production deploy. Across 89 commits from December 27 to January 1, Agxntsix consolidated 13 transfer tools into a single transferCall function, tuned transcription keyterm boosting to staff names and terms like RRSP and TFSA, and recorded 34 eval runs in a single day. A 12-destination production call checklist, dated December 31, verified every routing path before go-live.
The differentiator is the evals suite. Voice agents fail in ways unit tests never catch — a prompt tweak that fixes complaint handling can silently break executive gatekeeping. So Agxntsix built 13 scenario evals covering direct routing, ambiguous requests, CEO protection, complaint handling, frustration overrides, premature call endings, multilingual conversations, and cold callers, executed by a CLI runner (evals/run.ts) against the live Vapi assistant. No prompt change reaches OM Financial's lines without passing regression.
OM Financial went from business-hours-only human reception to a 24/7 multilingual AI receptionist on both office lines in five days — every routing rule version-controlled, every prompt change regression-tested, and every call producing eight structured analytics fields the firm never had before.
| Metric | Before | After |
|---|---|---|
| Call coverage | One human receptionist per office, business hours only | AI receptionist answers both office lines 24/7 |
| Languages served | Limited to languages on-duty staff happened to speak | 29 languages with seamless mid-conversation switching |
| Routing | Callers guessed names against an outdated website directory | Intent-based routing to 12 staff destinations within 1–2 clarifying questions |
| Executive access | CEO requests had no consistent screening path | Every CEO/president request routes through the executive assistant automatically |
| Call intelligence | No record of who called, why, or what happened | 8 structured analytics outputs per call written to Neon PostgreSQL |
| Quality assurance | Phone experience untestable except by calling in manually | 13 automated eval scenarios regression-test every prompt change |
Time-to-value was the headline: the discovery call and first production deploy happened the same calendar day — same-day deployment, five-day hardening, fully regression-tested.
OM Financial's AI receptionist identifies caller intent within one to two clarifying questions, then executes a warm transfer to one of 12 staff destinations — announcing the recipient's name and title, and passing a brief context summary so staff never answer cold. Transcription keyterm boosting on staff names and insurance terms keeps routing accurate.
Routing rules are explicit, not improvised. The system prompt encodes who handles what — advisor recruitment to the regional VP for insurance, RRSP and TFSA questions to the regional VP for investments, case status to new business, chargebacks to compliance, commissions to contracting and compensation — and a single consolidated transferCall tool executes the handoff. Deepgram Nova-3 keyterm boosting on every staff name and on vocabulary like RRSP, TFSA, and LLQP means names spoken with any accent still resolve correctly.
Agxntsix regression-tests every prompt change with an automated evals suite before it reaches OM Financial's live phone lines. Thirteen scenario files cover direct routing, ambiguous requests, executive gatekeeping, complaint handling, multilingual calls, and cold callers; a CLI runner executes them against the Vapi assistant and records scored results — 34 runs logged in one hardening day.
Each eval is a version-controlled JSON scenario — '02-ceo-protection' verifies that asking for the chief executive always routes to the executive assistant; '10-frustration-override' verifies an angry caller escalates correctly; '09-no-premature-endings' catches the agent hanging up too early. Combined with bidirectional sync scripts (sync:pull / sync:push) keeping git and the Vapi cloud config identical, every behavior change is tested, diffable, and reversible.
Yes. OM Financial's receptionist understands 29 languages through Deepgram Nova-3 multilingual transcription and switches languages mid-conversation when a caller does. A Google Gemini fallback transcriber, smart endpointing tuned for multilingual speech, and a low-confidence transcription hook keep accuracy high for the firm's diverse caller base across the Greater Toronto Area.
Multilingual support was engineered in layers: Deepgram Nova-3 in full multilingual mode with a Google Gemini fallback transcriber, Vapi's smart endpointing enabled because multilingual speech trips fixed silence thresholds, and a low-confidence transcription hook catching uncertain transcripts before they cause a misroute. A caller can start in English, switch languages mid-sentence, and still reach the right regional vice president — at 2 a.m., when no human receptionist was ever available.
Phase 2 is underway: a client-facing analytics dashboard with roughly 110 planned features — call logs with audio playback, transcripts, KPI charts, and a follow-up action queue — built on the eight structured outputs already flowing into Neon PostgreSQL.
Agxntsix deployed OM Financial's voice AI receptionist to Fly.io on December 27, 2025 — the same day as the discovery call — and completed production hardening by December 31, five days later. The build totaled 89 commits, including automated eval testing, a 12-destination routing checklist, and live call analytics.
OM Financial's AI receptionist enforces an executive gatekeeper pattern: any request for the CEO or president routes automatically to the executive assistant instead of ringing the executive directly. The rule lives in the version-controlled system prompt and is verified by a dedicated CEO-protection eval scenario before any prompt change reaches production.
OM Financial's receptionist extracts eight structured outputs from every call — caller type, primary topic, transfer destination, resolution, sentiment, prospect readiness, follow-up flag, and a call summary — and writes them to a Neon PostgreSQL database with a 60-plus-column schema covering transcripts, recordings, latency, and per-call cost.
Yes. Agxntsix loaded a comprehensive OM Financial knowledge base — company overview, service offerings, advisor support, compliance procedures, and FAQs — behind a query tool powered by Google Gemini 2.5 Flash. The receptionist answers basic questions directly and transfers to a specialist only when a caller needs personalized or expert advice.
Nothing — calls keep flowing. Agxntsix engineered OM Financial's backend for resilience: analytics writes never block the conversation, and webhook processing is budgeted to finish well inside Vapi's response window. A backend or database problem never interrupts a caller mid-conversation, because call routing itself runs on Vapi's telephony infrastructure.
Agxntsix builds voice AI the way software should be built — version-controlled, regression-tested, live in days. Book a discovery call.