🔒 Restricted

This page is gated

The Vani OS architecture briefing — what's actually built, the reference diagrams, and the generalization path. If you're meant to see this, you already have the passphrase.

ARCHITECTURE BRIEFING· PRIVATE & CONFIDENTIAL· UPDATED SEPTEMBER 2026

Vani OS

One AI conversion engine, configured — not rebuilt — for every high‑volume enrollment business.

Vani AI · OpenWorker · a fully self‑hosted voice stack — one outcome‑based platform, already running in production, now generalized past university admissions to test‑prep, coaching, and any lead‑to‑phone‑call conversion floor. What actually changes per industry is the State Machine and the Counsellor Definition Language (CDL) — not a rebuild.

00

The thesis

The mechanics of “inquiry to enrolled” are the same whether the outcome is a university seat, a JEE batch, or a motorcycle booking. Score the pool, brief the closer, log every touch, chase the SLA.

This isn't a pitch tool with an AI feature bolted on — it's one outcome‑based platform. Vani AI already runs the intelligence loop (scoring, briefing, CRM sync) against automotive dealer‑network accounts today. OpenWorker already runs the background automation. None of that was built for a lecture hall — it was built for anyone who converts inbound interest into a paid outcome by phone.

What's new in this briefing is twofold: a voice channel we own outright end‑to‑end, and the two pieces that actually change per industry — the State Machine that drives the conversation's stages, and the Counsellor Definition Language (CDL) that defines what gets said and when. Swap those two, keep everything else. Sections 01–03 show what already exists and how it generalizes; 04–05 show the owned voice stack and how it ships; 06 closes on the commercial model in its vertical‑agnostic form.

01

What's already in production

Three components, one working system today — not a roadmap slide. Every row below maps to a file in the working repositories. There's no separate desktop "cockpit" product in this stack — the counsellor‑facing surface (hit lists, pre‑call briefings, live Sales Assist, pipeline stages) lives inside Vani AI itself, driven by the State Machine and CDL, not a standalone app.

ComponentWhat it doesEvidence
Vani AI
Vibe Prospecting		 Account & lead intelligence, hit‑list scoring, CRM sync, research agents (Vivek, Vigil, Vibe, Vidya) — plus the counsellor‑facing surface itself: contextual pre‑call briefings, live Sales Assist, pipeline stages, the daily floor rhythm. Next.js + Supabase, multi‑org. In active use across two operating entities; account data already spans automotive OEM dealer‑network prospects (Hero MotoCorp, Maruti Suzuki, Royal Enfield) alongside the university‑admissions design. src/lib/crm-push.ts, src/lib/usage.ts
HitListPanel.tsx, PitchAdvisorPanel.tsx,
FollowUpIntelPanel.tsx, PipelinePanel.tsx,
TheFloorPanel.tsx, CrmApprovalPanel.tsx
OpenWorker
Volt AI		 Always‑on AI coworker: 25+ tool connectors, approval‑gated writes, scheduled automations. Provider‑agnostic — any model, or fully local via Ollama. coworker/engine.py
coworker/connectors/
vaniai‑bridge MCP integration layer wiring OpenWorker into Vani AI's accounts, hit‑lists, signals, and WhatsApp messaging. mcp-server.js
02

Architecture

This is what's actually running, not a reference design for later. The new layer is the voice channel — self‑hosted end to end. Everything else below it is already running; everything above it is a State Machine + CDL config, not a rewrite.

FIG. 1 — DATA & CALL FLOW

Prospect's phone Counsellor Telephony — Exotel / Plivo Vani AI — counsellor console hit‑list · briefing · pipeline · floor SELF‑HOSTED VOICE — PIPECAT Silero — voice activity detection IndicWhisper — speech to text Unsloth LLM — 4‑bit, fine‑tuned Kokoro / IndicParler — text to speech transcript + outcome ▸ hit‑list · approvals ▾ Vani AI intelligence · hit‑list scoring · CRM sync OpenWorker coworker automation scheduled vaniai‑bridge — MCP Customer CRM LeadSquared · Zoho · HubSpot sync — reads always, writes human‑approved VERTICAL CONFIG — State Machine + Counsellor Definition Language (CDL), per industry
Fig. 1 — Two entry points (a ringing phone, a counsellor's own console — both inside the same platform, not a separate cockpit product) converge on Vani AI, which syncs to the customer's own CRM through vaniai‑bridge. OpenWorker runs alongside on a schedule. A vertical‑config layer underneath swaps the State Machine and CDL for whichever industry is running.
03

One engine, many verticals

Strip the labels off any of these businesses and the pipeline underneath is identical: five stages, one scoring step, one assignment, one SLA clock. What actually changes between them is the State Machine (which stage follows which, what triggers an escalation) and the CDL (what gets said at each stage) — not the engine underneath. OpenWorker's own personas/ extension point follows the same pattern on the automation side — the mechanism that lets it act as a solutions engineer today can carry an “enrollment coworker” persona tomorrow, no change to its core engine.

FIG. 2 — SAME SHAPE, DIFFERENT VOCABULARY

INQUIRY SCORED ASSIGNED ENGAGED CONVERTED UNIVERSITY ADMISSIONS Student inquiry Hit‑list rank Counsellor assigned Call + WhatsApp Enrolled TEST‑PREP / COACHING Lead — web / walk‑in Batch‑fit score Closer assigned Demo class + follow‑up Fee paid, seat booked AUTO DEALER NETWORK Showroom / web inquiry Purchase‑intent score Sales exec assigned Test drive + follow‑up Booking confirmed
Fig. 2 — Same five‑stage shape, different vocabulary. The highlighted row is the next config to ship, not a new build.
$11.6B
India test‑prep & coaching market, 2025 — trending toward an estimated $26B by FY30 on differing methodologies. Structurally the same calling‑floor problem this platform already fixes for admissions.
₹3,000Cr
PhysicsWallah FY25 revenue, +55% YoY, 48 lakh students enrolled. IPO'd Nov 2025; one named proof that the enrollment‑conversion vertical is investable at scale.
0rewrites
Vani AI's scoring and CRM‑sync engine already targets automotive dealer‑network prospects (Hero MotoCorp, Maruti Suzuki, Royal Enfield) in the same codebase — the State Machine and CDL were never university‑specific plumbing.

Career Launcher and similar coaching networks run the same franchise/center model as the 20–25 seat BPO floor this platform targets. Market and company figures: Technavio, India test preparation market, Indian Television, test prep market to FY30, Inc42, PhysicsWallah FY25. Figures are third‑party estimates, not internal projections.

04

Owning the wire

The original design leaned on Torbi.AI and cloud model APIs for the voice channel. That's history now, not a plan — the replacement owns every node in the pipeline — orchestration, hearing, thinking, and speaking — on hardware that sits inside the building, and it's the box actually taking real calls today, not a diagram.

FIG. 3 — CLOUD‑DEPENDENT vs. OWNED

BEFORE Phone call Telephony Cloud STT API Cloud LLM API Cloud TTS API $ per token/min · +200–400 ms per hop · data leaves the network OWNED — LIVE TODAY Phone call Telephony ONE ON‑PREM GPU BOX Silero VAD IndicWhisper STT Unsloth LLM, 4‑bit Kokoro TTS ₹0 marginal / call load-tested to N=5 never leaves the building
Fig. 3 — Three metered cloud hops become one box on a fixed cost. Pipecat (orchestration, BSD‑2) stays the same either way — only what's behind it changes.

Telephony still runs through Exotel or Plivo — both licensed for Indian PSTN, both publish Pipecat integrations, at roughly ₹0.60/minute outbound. That's the only per‑call cost left in the system; everything behind it is owned.

N=5
Real concurrent-call load test against the live box (2026‑09‑05), real audio through the real /api/start-call path, not synthetic. Default Whisper STT degrades under load — time-to-first-transcription climbs 2.9s→5.6s from N=1 to N=5. Swapping in Deepgram STT (already a real code path, just needed its credential wired through) holds ≤1.5s flat across the same range — the fix is a provider swap, not new hardware.
0cloud VMs
LiveKit — call transport — is now self-hosted on this same box (native Docker inside WSL2, not Docker Desktop), reachable through a Cloudflare Tunnel with TURN-over-TLS. The GCP VM that used to carry this hop is gone — the account was fully decommissioned 2026‑09‑07, zero billed compute remaining.
$3,499
NVIDIA Jetson AGX Orin 64GB Developer Kit — the actual next hardware tier, sized and priced for real: 275 TOPS, 204.8 GB/s memory bandwidth, a reasoned 4–8 concurrent-call estimate scaled from this box's own component benchmarks. Not yet load-tested itself — see §05.
05

Deployment path

Same containers, three real tiers — the first is live, not a plan.

LIVE — PRODUCTION

Lenovo (native Docker, WSL2)

  • GTX 1650 Ti, 4GB VRAM — LLM pinned to GPU
  • Deepgram STT (not Whisper) — the real fix for §04's load-test finding
  • LiveKit self-hosted here too, via Cloudflare Tunnel + TURN‑over‑TLS — no GCP hop
  • Real-load-tested to 5 concurrent calls (2026‑09‑05); 2 confirmed in real production use
NEXT

AGX Orin 64GB — pilot box

  • NVIDIA Jetson AGX Orin 64GB Developer Kit — $3,499, 275 TOPS, 204.8 GB/s bandwidth
  • Kcube Labs-owned, shipped to a pilot university's site — zero hardware CapEx to them
  • 4–8 concurrent calls, reasoned from component benchmarks — not yet load-tested itself
TARGET

Growth box / hybrid cloud

  • RTX 4090‑class tower (~10–12 concurrent) or full LiveKit SFU + cloud STT/LLM/TTS
  • Exotel / Plivo trunk for real PSTN traffic either way
  • Same Docker images — copy, not rebuild
06

Commercial model, generalized

The outcome‑fee structure built for university admissions already generalizes without changing its logic — only the label on the outcome event changes.

VerticalBaseline metricOutcome event
University admissions Same‑period‑last‑year admits, matched cohort Enrolled student, above baseline
Test‑prep / coaching Trailing‑term seat bookings Fee paid, batch seat confirmed
Auto dealer network Trailing‑quarter bookings per showroom Booking confirmed, above baseline

Fee is payable only above a baseline locked in writing before the campaign starts, and only for a student, learner, or booking with at least one logged Vani OS touch — audited the same way regardless of vertical.