Voice is the most natural way to talk to software, and real-time models are now fast enough to hold a proper conversation. If your product would work better with users speaking instead of typing, such as rehearsing a sales call, being coached, filling in a form hands-free or asking a question while their hands are busy, a voice AI agent can do it in the browser or inside your app. I'm Md. Emamul Mursalin, a freelance AI developer, and I build real-time voice features at a fixed price for teams in the US, UK, Saudi Arabia and the Gulf.
The demo is the easy part. Production is latency, cost and data: replies that start within a few hundred milliseconds, a bill that does not grow faster than your revenue, and conversations that never leak between customers. Those are the three things I design for first.
Md. Emamul Mursalin · Rajshahi, Bangladesh · remote worldwide · UTC+6 · Updated
A voice agent streams the user's speech to a real-time model, which answers out loud mid-conversation and copes when the user interrupts. Around that conversation sit the parts that make it useful: the persona or role the agent plays, the instructions and knowledge it works from, the tools it may call, and a transcript you can score, search or hand to a person.
I build voice into web apps and into React Native apps for iOS and Android. This page is about voice inside your own product, where the user talks to your app directly, rather than agents that answer phone lines.
Adding voice to a product you already have is my AI Integration work. One voice feature, such as a spoken assistant or voice input for an existing workflow, is the Starter package: $799 and one to two weeks, with prompt design and test cases, integration into your codebase, model keys kept on the server with cost caps, usage logging and a handoff. When the agent has to answer from your own documents or data, or run workflows with review steps, Professional is $2,499 over three to five weeks.
A new product built around voice, such as a coaching, practice or assistant app your customers sign up for, is an AI Agent SaaS build. Starter is $1,499 over two to three weeks for a single-purpose product with sign-in, Stripe billing and usage metering. Professional is $4,999 over five to eight weeks for a multi-tenant product with team workspaces, tiered billing, LLM observability and an admin panel.
Real-time voice models are billed by audio usage, so every proposal includes a running-cost estimate from your expected minutes, and each build ships with per-user limits and a hard monthly cap. For work outside a package I bill $20 to $30 an hour.
Virtual Client is a B2B coaching platform where teams rehearse live conversations against AI personas and get rubric-scored feedback. I was the lead engineer on a team of five, building it for Robbie Boyle. The personas run on Gemini Live for two-way voice. A Vercel Edge Function proxies the audio stream so it never round-trips through the backend, which keeps latency under 300 milliseconds for 95% of sessions, and a token bucket limits concurrent streams per organisation to keep costs under control.
After each session, OpenAI and Anthropic models score the conversation against a rubric. The platform runs across several organisations in production with roughly 2,500 practice sessions delivered, and the isolation between organisations is guarded by more than 370 security tests, so no organisation can see another's sessions.
Most of the work is in the plumbing around the model: where the audio goes, who is allowed to start a session, how long it may run and what happens to the transcript afterwards.
I am based in Bangladesh at UTC+6 all year, so a UK team gets me for its whole afternoon, a US East Coast team gets its morning and its evening, and a Saudi team shares my Sunday-to-Thursday week. Code lives in your repository, every change goes through a pull request your team can review, all IP is assigned to you on final payment, and payment is by milestone in USD through Wise, Payoneer, bank transfer or a Stripe invoice.
Proof
Projects built with the same stack and patterns.
Robbie Boyle · EdTech / Corporate Training
Teams rehearse live conversations against AI personas and get rubric-scored feedback.
“This project was handled in an excellent way. The seller understood the task clearly. Progress updates were helpful. The final output was accurate and polished. Everything arrived as promised. Glad to recommend this seller.”
Alexi Denise
Individual client
“He successfully completed all my tasks quickly and with excellent quality.”
Imran Mohammed
“Great work. Great communication, and a quick turnaround. I'd definitely order again.”
Agrossen
Fixed price
Fixed price, fixed scope. Starter and Enterprise tiers are on the pricing page.
RAG over your data plus automated workflows
$2,499/project
Launch offer$1,62435% off for the first 5 clients
3-5 weeks
Multi-tenant AI SaaS with RAG and agent workflows
$4,999/project
Launch offer$3,24935% off for the first 5 clients
5-8 weeks
Full-featured AI app with subscriptions and analytics
$3,999/project
Launch offer$2,59935% off for the first 5 clients
6-9 weeks
Case studies with the same stack and patterns.
Costs, architecture and how-to guides from the blog.
Questions & answers
With me, $799 to add one voice feature to a product you already have, or $2,499 when it must answer from your own data or run workflows. A new product built around voice starts at $1,499, and $4,999 for a multi-tenant version with team workspaces and tiered billing. Voice model usage is billed separately and estimated in the proposal.
Fast enough to feel like a conversation. On Virtual Client, streaming audio straight to Gemini Live through an edge function keeps latency under 300 milliseconds for 95% of sessions.
Gemini Live runs the real-time voice on Virtual Client in production, with OpenAI and Anthropic models scoring each session. For a new product I choose the model on latency, voice quality and cost per minute, and keep it behind one interface so it can be swapped later.
Yes. Voice runs in the browser or in a React Native app for iOS and Android, and voice input is one of the AI features included from the Starter tier of my AI-powered mobile app packages.
Not as a standard package. My voice work is inside web and mobile products, where the user talks to your app directly. If you need an agent on a phone line, tell me on the first call and I will say honestly whether it is a good fit.
Model keys stay on the server, transcripts are stored in your own accounts, and when the product serves more than one organisation the isolation between them is enforced and tested in CI, as it is on Virtual Client.
Next step
A 30-minute call is enough to tell you which package fits, what I would cut, and what it will cost to run. No pitch, no obligation.
Or email hello@mursalinsdesk.com
