Deterministic tool orchestration
Patent-pending orchestration engine gates every tool call against a state-machine graph of valid next-state transitions. The LLM suggests; the engine decides. Eliminates an entire class of LLM-hallucinated actions.
Trade-ops AI
AI that runs the job — not just answers the phone.
Vaxio builds AI agents that run the trade-ops workflow end-to-end — booking, estimating, dispatching, invoicing, follow-up — with the owner in the loop on the sensitive actions. Frontbell, our launching product, brings it to service trades: a voice-first agent that runs the front-desk workflow at sub-second latency, drafts the estimate, queues the dispatch, and escalates anything outside its configured authority. State-machine gating means the agent can't fake a booking, hallucinate an estimate, or take an action the owner didn't approve.
The opportunity
Service trades miss the calls that build their business.
Painters, cleaners, and small remodelers run on inbound demand they can't capture. Crews are on roofs, in attics, behind ladders. The phone rings — and a competitor wins the job.
62%
of calls to small service businesses go unanswered or to voicemail
Salesforce SMB Service Report
$1,200
average lifetime value of a single missed lead in residential trades
Industry composite, 2025
$650B
US trades industry spend serviceable by software — and the category's largest SaaS player has under 10% penetration today
ServiceTitan S-1 (Dec 2024)
Frontbell answers every call, qualifies the work, books it on the calendar, and texts back confirmation — without the owner picking up the phone.
Generic agent frameworks don't survive contact with deterministic orchestration. We built our own.
Generic LLM chat platforms can't hold a phone conversation, gate tool calls against a state machine, and execute a real business workflow end-to-end. Latency, interruption, tool-calling reliability, and action honesty break simultaneously. We built Frontbell on a custom streaming voice pipeline AND a patent-pending state-machine tool-execution engine — so the agent doesn't just sound human, it executes the workflow deterministically.
Persistent workflow context
Conversation memory across phone, SMS, and chat. Each interaction builds on the last — no repeated names, no repeated addresses, no dropped workflow state.
Sub-second voice latency
p50 ≤ 800ms first-token, p95 < 1.2s (production target). Streaming ASR, streaming LLM, streaming TTS — no batch handoffs. How the agent feels human while doing the work.
Pluggable model backends
Frontier LLMs for reasoning. Smaller models for intent and routing on the path to self-hosted inference. We're vendor-flexible, not vendor-locked.
Full-duplex orchestrator
Custom WebRTC pipeline with local VAD and RMS-based echo cancellation. Callers can interrupt mid-sentence and the agent yields, like a person would.
Carrier-grade telephony
Twilio Subaccount-isolated tenants, 10DLC-compliant SMS, per-tenant brand registration. Not a Zapier wrapper around someone else's stack.
What we ship
Product
Engine
Engine
Frontbell — and the engines we built underneath to make it work.
One product, three trades at launch. Behind it: the engines we built to make conversational software shippable to small businesses.
Frontbell
Real-time voice agent that answers every call, books jobs, and confirms appointments. Painting, cleaning, and carpentry/small-remodels at launch.
frontbell.ai · public launch mid-July 2026
CE-OS
Conversational Execution engine. State-machine tool orchestration, intent classification, persistent memory. Powers Frontbell. Patent-pending (USPTO).
FUSE
Integration layer. Connects CE-OS to calendars, SMS, calling, payments, and CRMs without per-customer custom integration work.
Ready to hear it?
See Frontbell in action — or talk to us about bringing trade-ops AI to your business.