Skip to content
mahadizulfiker
All work
Web app2026SeloraX

SeloraX Voice AI

Bangla AI call agent, grown into a multi-tenant platform merchants run themselves.

Primary authorPrimary builder of the voice-AI engine.

A working AI agent that answers and places calls in fluent Bangla end-to-end, now running as a multi-tenant product: merchants onboard themselves, buy a plan, and manage their own agents and recordings across 11 console pages. (Shown as an architecture narrative, client system, no code or data exposed.)

A voice-AI call agent for Bengali-speaking merchants: an AI that holds real phone conversations in Bangla over a SIP gateway. Built the voice engine first, then the platform around it. Each merchant is an isolated tenant with a self-serve console for agents, campaigns, and call recordings, reached through a public merchant API and a front door with real pricing. An entitlement gate ties a plan to what a tenant may actually run, inbound calls route by DID, and per-call cost is recorded against the call that caused it so the economics stay legible.

SeloraX Voice AI call pathA phone caller reaches a SIP gateway, which bridges into a realtime LiveKit room feeding Gemini Live. The reply is spoken by a custom Bangla text-to-speech voice and returned down the same call, the latency-critical path.CALLERinbound / outboundSIP GATEWAYtelephonyLIVEKITrealtime audioGEMINI LIVEreasoningBANGLA TTScustom voicespoken reply
call path · accent marks the latency-critical return

This is a client system, shown as an architecture narrative only: no source code, credentials, or user data are exposed.

Problem

Bengali merchants needed an automated phone agent, but off-the-shelf voice AI has no usable real-time Bangla speech, latency and pronunciation both break the illusion of a real call.

Constraint

A phone call tolerates only a few hundred milliseconds of round-trip lag before it feels broken, and the voice had to sound natural in Bangla, not machine-read.

Decision

Bridged a SIP telephony gateway into a realtime LiveKit audio pipeline feeding Gemini Live, and paired it with a custom-trained Bangla TTS voice tuned for phone-band audio.

Tradeoff

A custom TTS voice is far more work than a hosted one and has to be maintained, but no hosted option produced acceptable conversational Bangla.

Result

A working AI agent that answers and places calls in fluent Bangla end-to-end, now running as a multi-tenant product: merchants onboard themselves, buy a plan, and manage their own agents and recordings across 11 console pages. (Shown as an architecture narrative, client system, no code or data exposed.)

next step

Open to full-time engineering roles.

I reply within a day.

Next project

Smart Sheba
Chat on WhatsApp