Production voice AI agents that ship to real phone numbers.
I build multi-tenant voice AI platforms — real-time STT → LLM → TTS streaming, Twilio Programmable Voice, ElevenLabs, OpenAI Agents, knowledge-base grounding, credit wallets, Stripe metered billing. The hard parts of voice AI aren't the prompts. They're latency, billing, and tenant isolation. I've shipped all three in production.
- Sub-second end-to-end latency on the full STT → LLM → TTS loop
- Multi-tenant by design — RLS-enforced isolation, per-tenant tools
- Barge-in handling, calendar-aware agents, and metered billing built in
Who I build voice AI for
Voice AI is the 2026 freelance category growing fastest, and it's also the one most teams underestimate. The buyers below have shipped something to the demo stage and hit a wall on the production parts.
AI voice startups past prototype
Your voice agent demo works in a happy-path screen recording. Real production needs sub-second latency, interruption handling, multi-tenant isolation, billing, and a way to debug bad calls. That's the gap I close.
Sales-tech and customer-experience teams
You want voice agents for outbound sales, inbound support, or appointment scheduling. They need to integrate with your CRM, your calendar, your billing — and not embarrass you on a real call.
Agencies with voice-AI clients
Your agency landed a voice AI deliverable but the dev team has only shipped chat. Bring me in white-label to deliver the voice production engineering — I work under your brand, in your client's Slack.
Operators replacing call-center work
You run a real business — recruitment screening, hospitality booking, medical intake — where calls are a real cost line. You want a calendar-aware, knowledge-grounded agent that handles the 80% of calls that don't need a human.
Why most voice AI builds stall at production
The chat-AI → voice-AI gap is bigger than it looks. A chat agent that responds in three seconds feels snappy. A voice agent that responds in three seconds feels broken — the silence is unbearable on a phone call. The whole architecture has to be rebuilt around streaming: partial-token forwarding from LLM to TTS, chunked playback, aggressive caching, all timed to keep the end-to-end loop under a second.
Three other things kill most voice builds. First, no real interruption handling — the agent talks over the user, or the user can't interrupt mid-sentence without crashing the session. Second, no multi-tenant isolation — agents leak knowledge, billing breaks, prompt injections from one tenant affect another. Third, no observability — when a call goes wrong, nobody can figure out which leg of the pipeline broke, so the same bug repeats every Tuesday.
I've built and shipped voice AI in production — a multi-tenant voice platform on Twilio + ElevenLabs + OpenAI Agents, with credit wallets, Stripe metered billing, calendar-aware agents, real-time streaming. The same architectural patterns apply to most builds: streaming STT, function-calling LLMs, chunked TTS playback, barge-in state machines, per-tenant memory and tool isolation, metered billing per call-second. The architecture is well-understood. Most teams just haven't put one in production before.
How I build production voice AI
Every voice agent I ship follows the same architectural pattern, refined across production builds and several private engagements.
Streaming-first architecture
STT streams partial transcripts to the LLM while the user is still talking. The LLM streams tokens to TTS as they generate. TTS streams audio chunks to Twilio as they synthesize. End-to-end latency under 900ms in good network conditions — the user never sits in silence.
Interruption + barge-in handling
Users can interrupt the agent mid-sentence; the agent stops, listens, and resumes appropriately. Built on Twilio Media Streams with a state machine that tracks who's talking — not the brittle 'just pause TTS' approach most demos use.
Multi-tenant by design
Every customer gets isolated agents, isolated tool access, isolated knowledge bases, isolated billing. Postgres RLS, per-tenant API keys, per-tenant prompts. Zero cross-tenant leakage — verified by audit log, not assumed by hope.
Function calling + tool isolation
Agents call your APIs to look up customers, book appointments, check inventory, transfer calls. Per-tenant tool registration, schema validation, rate limiting, audit logging. Bad function calls are caught and recovered — not crashed.
Knowledge base ingestion
Per-tenant document ingestion (PDFs, websites, FAQ files) → embeddings → pgvector → RAG context injected into the system prompt. Updates without rebuilding the agent. Citations surfaced in the post-call transcript so customers can verify what the agent said.
Metered billing + credit wallets
Stripe metered billing per call-second. Credit wallets, top-ups, plan upgrades, usage alerts. Every minute is tracked, billed, and exportable. Customers self-serve plan changes without your team touching the database.
Engagement shape
Voice AI projects ship most predictably as scoped milestone-based builds. We start with a working call to a single phone number, then layer in tenancy, billing, and operations.
Discovery (Day 0)
60-minute call. You demo what you have. I walk through your latency budget, tenancy model, call-volume projections, and billing requirements. NDA signed before we go deeper if needed.
Architecture + scope (Week 1)
Written architecture doc: streaming pipeline diagram, tenancy model, billing model, tool integration plan, observability plan, fixed-price quote with milestone breakdown. Full picture before any code.
Voice MVP (Weeks 2–4)
Single tenant, single agent, single Twilio number. Real streaming. Real interruption handling. Real tool calls. By end of week 4 you can call the number and have a conversation.
Production hardening (Weeks 5–8)
Multi-tenant isolation, Stripe metered billing, knowledge base ingestion, admin dashboard, observability, error recovery. Production-ready, not just demo-ready.
Launch + handoff (Week 9+)
Real customer cutover, monitoring, runbooks, on-call playbook, Loom walkthroughs. 30 days of post-launch support. Optional retainer for continued iteration.
What you get
Investment
Two engagement shapes — scoped MVP build or full multi-tenant production. Fixed-price, milestone-based, no hourly surprises.
Voice MVP
$8,000 – $18,000
3–4 weeks. Single tenant, single agent, real Twilio number, real streaming. Right for validating the use case with real users before going multi-tenant.
- Single-tenant voice agent
- Real-time STT → LLM → TTS streaming
- Interruption + barge-in handling
- 1–2 tool integrations
- Basic admin UI
- Call recording + transcripts
- 30 days post-launch support
Production Voice SaaS
$25,000 – $65,000
8–10 weeks. Multi-tenant production build with per-tenant isolation, Stripe metered billing, knowledge base ingestion, full admin dashboard. Right for funded teams shipping a voice AI product.
- Multi-tenant agent platform
- Per-tenant RLS isolation
- Stripe metered billing + credit wallets
- Knowledge base ingestion + RAG
- Function calling + tool registration
- Full admin dashboard
- Observability + alerting
- Architecture docs + handoff
- 30 days post-launch support
Voice AI Retainer
From $10,000/month
Embedded engineering for ongoing voice AI work — new agents, new integrations, latency optimization, scale. Common pattern post-launch.
- Dedicated weekly capacity
- Architecture ownership
- Priority response on production issues
- New agent + tool development
Twilio + ElevenLabs + OpenAI usage costs run separately — typically $0.06–$0.15 per call-minute depending on model selection. All service accounts in your name.
Tech Stack
Working With Clients Across
Recent Work
Related case studies
Frequently asked
How fast is the end-to-end latency?
Sub-second on a typical call in good network conditions. We measure ~150ms to first STT partial, ~250ms to first LLM token, ~150ms to first TTS chunk, total round-trip under 900ms before the user hears the first phoneme of the response. The architecture is built around streaming, not request-response — that's the difference between a snappy agent and one that feels broken.
Can the agent handle interruptions and barge-in?
Yes. Twilio Media Streams gives us real-time bidirectional audio, and we run a state machine that tracks who's talking. User starts speaking → TTS playback pauses → STT resumes → new turn begins. Built and tested in production.
How do you handle multi-tenant isolation?
Postgres Row-Level Security on every table that touches customer data. Per-tenant API keys, per-tenant prompts, per-tenant knowledge bases, per-tenant tool registrations. Cross-tenant leakage is impossible by architecture — not by hoping the prompt won't get jailbroken. Audit logs prove it.
Do you integrate with our CRM, calendar, or database?
Yes. Function calling exposes your APIs to the agent — HubSpot, Salesforce, Attio, Google Calendar, Outlook, custom REST. Each tool is registered per-tenant with schema validation and rate limiting. The agent can read customer data, book appointments, transfer calls, log notes — whatever you expose, scoped to what each tenant should see.
How is billing handled?
Both Stripe metered billing per call-second and credit wallets, depending on the product model. Plan upgrades, downgrades, top-ups, and usage alerts are all in the admin dashboard. I run the credit-wallet pattern in production; either pattern is straightforward to ship.
Can you do this for non-English languages?
Yes. Whisper handles 50+ languages for STT. ElevenLabs supports 29 languages for TTS with voice cloning. OpenAI and Claude both handle non-English content well. Multi-language is a per-tenant config — tune prompts and voices per language.
What happens if Twilio, OpenAI, or ElevenLabs has an outage?
Each leg has a fallback. STT can fall back from Whisper to Twilio's native STT. LLM can fall back from OpenAI to Anthropic. TTS can fall back from ElevenLabs to Twilio's native voices. The agent degrades gracefully rather than dropping the call.
Do you sign NDAs and enterprise procurement paperwork?
Yes. Standard mutual NDAs before discovery. MSAs + SOWs for enterprise builds. SOC 2 / HIPAA-aware architecture available on request — voice AI in regulated industries needs private endpoints, audit trails, and PII handling built in from day one, not retrofitted.
Ready to start?
Every engagement starts with a 30-minute discovery call. I'll listen, ask sharp questions, and send a written proposal within 48 hours.
Explore Other Services