Answers the phone
Connects to the Yeastar or SIP system the business already runs, so calls reach the agent on the numbers customers have always dialled.
AI voice agent case study
Vocale
An agent that answers a business's phone line, finds the answer in that business's own documents, and opens the support ticket before the caller has hung up, on the numbers they already own, and on their hardware if they want it there.
- IndustryInbound support and telephony
- UsersSupport teams on a Yeastar or SIP PBX
- CoverageEvery call, around the clock
- ServicesProduct Design, Development, AI
Overview
Vocale puts an AI agent on a phone line a business already has. A customer dials the number they have always dialled; the agent picks up on the first ring, speaks their language, answers from the company's own manuals and policies, and when the answer is really "somebody needs to look at this" files a support ticket with the details already collected.
The decision that makes or breaks the economics is where the models run, so the platform ships two engines behind one product. On the hosted tier, speech, reasoning and voice come from a third-party API and every minute is metered. On the self-hosted tier all three run on a GPU under the business's own control, no audio or document leaves the building, and a call costs electricity. Switching between them is configuration; the agent behaves the same either way.
Video
- Next.js 16
- TypeScript
- Tailwind CSS
- Supabase
- PostgreSQL
- LiveKit
- WebRTC
- SIP telephony
- OpenAI gpt-realtime
- Python
- FastAPI
- Redis
- Pinecone
- Docker
- Next.js 16
- TypeScript
- Tailwind CSS
- Supabase
- PostgreSQL
- LiveKit
- WebRTC
- SIP telephony
- OpenAI gpt-realtime
- Python
- FastAPI
- Redis
- Pinecone
- Docker
Key capabilities
Grounded in your documents
Manuals, policies and price lists are uploaded, indexed and searched at the moment of the question so the answer comes from the business's own material rather than the model's memory.
Opens tickets mid-call
When a caller needs a person, the agent gathers the subject, the contact and the urgency while it still has them on the line, then files the ticket itself.
Knows who is calling
An existing customer database connects read-only, so an inbound number is matched to an account before the agent says a word.
Two engines, one product
The same agent runs on a hosted model API or entirely on a private GPU. The business picks on cost and on where its data is allowed to go.
A support line is only as available as the people answering it, and the questions arriving on it are overwhelmingly the same handful: what are your hours, is this covered, how do I reset it? The answers are already written down in a manual, a policy, a price list, just not anywhere a caller can reach, and an operator can find while somebody waits. Putting an AI on the line fixes the availability half of that. Renting one by the minute quietly replaces it with a cost problem, because the better it works, the more it is used, and the more it costs each month.
Calls outside office hours
Evenings, weekends and holidays go to voicemail and what lands there is mostly the routine enquiry a person would have closed in a minute.
Answers nobody can reach
The information exists in documents, but not in any form somebody can consult with a caller waiting on the line.
Priced by the minute
Managed voice-AI platforms bill per call minute, so the bill grows with adoption the wrong incentive to put on a support line.
Vocale sits between a business's phone system and its knowledge. The PBX connects from a form, and the routing is provisioned behind it. Documents are uploaded and indexed. The agent's manner is set in plain language rather than in code, and the whole thing can be rehearsed in a browser before a real call reaches it. Once it is live, everything it did is on the record: the transcript, the searches it ran, the tickets it opened, so the team can judge whether it is doing the job before trusting it with more of one.
Connects to the PBX, not around it
Yeastar and compatible SIP systems register from a form, and the business keeps its provider, its numbers and its call plan.
Retrieval before it speaks
A question that needs a fact triggers a search across the indexed documents, and the reply is assembled from what came back, with nothing found being an answer the agent is allowed to give.
Hosted or self-hosted, same behaviour
Speech, language and voice run behind a third-party API or on a private GPU. Moving between them is configuration, not a rebuild.
UX researchFeb–Mar
MVPApr–Jul
OperationAug–Sep
DiscoveryFeb
ArchitectureMar
UI designApr–May
Backend and local AIMay–Jul
TestingJul–Aug
OperationAug–Sep
Test it before a customer does
A sandbox that runs the whole pipeline in the browser. Pick a caller, start the call, and the transcript arrives live beside it as the agent hears and answers. Every session it has ever handled, browser or phone, inbound or outbound, is one click away in the rail on the left.
The documents it answers from
Files are dropped straight onto the page, then parsed, split and indexed. Each row reports how many pages went in and how many searchable pieces came out, so it is obvious at a glance when a document is ready to be quoted on a call and when it is still being worked on.
What it could not answer becomes a ticket
The agent collects the subject, the customer and the urgency while the caller is still on the line and files the ticket itself, transcript attached. Support opens the queue to a list of real problems in the caller's own words instead of a column of missed calls.
It knows who is calling
An existing customer database connects read-only, with a mapping step for whichever columns hold the name and the number. From then on an inbound call is matched against it before the agent speaks, and the records stay where they are. Vocale reads them; it does not become the place they live.
Numbers, assigned from the console
Telephony is the one part a business should not have to get right alone, so provisioning is an operator job rather than a self-service form. Each number is registered, pointed at an organisation, and switched on from here, which doubles as the fastest way to see which tenants are actually taking calls.
What every minute costs
Spend is metered the way the models bill it: caller speech, agent speech, and replayed history, broken out per model and per organisation. It is what turns the choice between a hosted API and a private GPU into a decision with a number attached rather than an argument.
What matters on a support line is not accuracy in the abstract; it is whether the call was answered and whether the answer was right. Vocale answers all of them, at any hour, in the caller's language, and leaves a transcript behind so the second question can be settled. The cost below is what the platform metered on the hosted tier over a month of real calls; the self-hosted tier moves that line onto hardware the business already pays for, and the per-minute figure stops moving at all.
- 24/7
- Call coverage, with no rota and no queue
- $0.06
- Metered cost per call minute, hosted tier
- 6
- Languages the agent can be switched to
- 2
- Ways to run it: hosted API or private GPU
User goal
- Caller dials in
- The agent answers
- Looks up the customer
- Searches the documents
- Opens a support ticket
- Emails it
Interface & channels
- Yeastar PBXPhone calls, bridged into LiveKit over SIP
- livekit-clientVoice testing in the browser
- NodemailerTicket emails, sent over SMTP
- User and admin portals
- Next.js 16
- React 19.2
- React Compiler
- TypeScript 5
- Tailwind CSS v4
- shadcn/ui
- Radix UI
- Framer Motion
- TanStack Query v5
- React Hook Form
- Zod v4
- EN / IT
- next-themes
- sonner
- cmdk
- lucide
- react-pdf
- md-editor
- lamejs
- pnpm
Planning & orchestration
- LiveKit Agents 1.6The agent worker behind every call
- Ticket flowA state machine of our own, from first question to sent ticket
- PydanticTyped orchestration models
- Silero VADTurn-taking, with LiveKit's turn detector
- Session gateCall caps per organisation and across the system
- StreamlitA scenario lab, for testing only
Memory & context
- RedisShort-term memory: a sliding window of the last 15 turns
- SupabaseLong-term memory: every turn of every conversation
- Tenant presetsConfiguration and presets for each organisation
- RAG plane
- Supabase StorageThe source documents
- pymupdf4llmPDF to Markdown
- ChunkerSplit on headers, then about 500 tokens with 50 overlapping
- multilingual-e5-largeEmbeddings, through Pinecone Inference
- PineconeA serverless vector index
- bge-reranker-v2-m3Reranks to the best 5, above a 0.20 threshold
- RedisIndexing jobs and their progress
- FastAPIThe retrieval API, served by Uvicorn
Models
- gpt-realtime-2.1Speech to speech: listening, reasoning and speaking in one model
- gpt-realtime-2.1-miniThe lighter realtime model
- gpt-4o-mini-transcribeTranscripts of what the caller said
- Legacy GPU pipeline, kept for rollback
- faster-whisperSpeech to text
- vLLMServing Qwen2.5-VL-7B-AWQ
- Piper and Kokoro-82MText to speech
Tools & actions
Functions the agent calls while the caller is still on the line.
- search_knowledge_baseAnswers from the documents, through RAG
- lookup_customer_detailsFinds the caller in the customer's own MySQL
- search_customer_identityConfirms who is calling
- Ticket toolsCreate, update and submit the ticket
- update_ticket_draftKeeps the draft and the contact details current
- end_callEnds the call cleanly
- Call recordingPCM to WAV, with FFmpeg
State & data
- PostgreSQL 17.6Supabase, self-hosted in Docker, across 47 migrations
- SupabaseAuth, and storage for documents and recordings
- KongThe API gateway, with Supabase Studio
- Redis 7Three instances: LiveKit, agent memory and RAG jobs
- PineconeThe vectors
- MySQLEach customer's own database, outside the system
- Customer Data PlaneFastAPI, with a pooled PyMySQL connection to each customer
Observability, safety & evals
- Cost meteringCost per call and per organisation, from usage and model pricing
- Resilience metricsSelf-repeats, talk-over, false barge-ins and latency
- Admin dashboardsMonitor, agent state, agent activity and database logs
- journaldFile logs, with systemd alerts on failure
- pytest26 test modules, a replay harness, and a transcript judge across 5 regression classes
- Call scenariosWritten in YAML, and replayed against the agent
- Agent safetyA leaked-reasoning detector, privacy and email read-back rules, an inactivity watchdog
Security
- NextAuth v5Credentials sign-in, bcrypt hashes, a 24-hour JWT in a secure cookie, email verification
- Access controlRoute guards, admin authorisation, and row-level security on Supabase
- AES-256-GCMCustomer database passwords encrypted at rest, readable from Node and Python
- Service keysAPI keys between services, and a secret for scheduled jobs
- Rate limitingOn every sign-in route
- Security headersCSP, HSTS with preload, no framing, and the microphone for this site only
- Private networkThe calling stack reachable only over VPN, and an allowlist on the SIP trunk
- WireGuardOne outbound IP for customers to allow into their MySQL
Deployment
- Ubuntu 24.04One self-hosted server, behind the VPN
- NGINXTLS, and one domain for both portals and Supabase
- systemdBoth portals, the agent worker, and a ticket email timer every 2 minutes
- Docker ComposeSupabase, LiveKit and SIP, RAG and customer data, on one shared network
- supervisordThe media plane container
- Vast.aiThe GPU box from the vLLM era
What a business can do from the dashboard
Seven areas behind one sign-in. Telephony and knowledge are set up once; operations, ticketing and customer records are the day-to-day. Anatomy of a single spoken turn
Speech in, speech out, in one pass — with both halves written down as they happen. On the self-hosted tier every box inside the agent is a model running on the business's own GPU.
Being interrupted mid-sentence
Solution:Callers talk over an answer the moment they have understood it, and an agent that keeps going sounds broken. Voice activity detection cuts the outgoing audio as soon as the caller speaks, then truncates the agent's own transcript to what was actually heard — so the conversation continues from the call the caller experienced rather than the one the system had planned.
Silence while it thinks
Solution:Searching the documents takes long enough to read as a dropped call. The agent fires a short, natural filler half a second in and cancels it if the search returns first, which buys retrieval its time without any of it landing as dead air.
Text that does not read aloud
Solution:Written for a screen is not written for a voice: currency amounts, dates as digits, acronyms. A normalisation layer rewrites amounts, dates, times and abbreviations into their spoken form, per language, before they ever reach the speech model.
One GPU, several conversations
Solution:A live voice session holds its models in video memory, so capacity is bounded by VRAM rather than by raw compute. Sessions are budgeted per card and the workload is pinned across cards explicitly — left to default pooling, speech and language work land on the same GPU and starve each other.
Questions
No. Vocale connects to the phone system you already run — Yeastar and compatible SIP platforms — and calls keep arriving on the numbers your customers already have. The connection is made once from the dashboard, and the trunk and routing are provisioned behind it.
Whichever way suits you. The same agent runs on a hosted model API, which is the quickest way to start, or entirely on a GPU under your control, where no audio, transcript or document leaves your infrastructure. It is a configuration choice, and it does not change what the agent can do.
It answers from your documents rather than from memory. A question that needs a fact triggers a search across everything you have uploaded and the reply is built from the passages that come back. When nothing relevant is found, saying so and offering to open a ticket is the correct answer, and the one it gives.
It hands over. The agent collects the subject, the contact and the priority while the caller is still on the line, files a ticket in your queue and attaches the transcript — so whoever picks it up starts from what was actually said rather than from a callback request.
Losing calls after five o'clock?
Let's put an agent on your line.