CypherLM
Pakistan's First No-Code, Voice-AI Agent Platform. Build. Deploy. Speak. — No engineers required.
Context
The Challenge
In Pakistan's growing digital economy, thousands of SMBs pour money into ads to drive traffic, but when a lead lands on their site at 2 AM, no one answers. Existing AI voice solutions are priced in USD, require deep API knowledge, and lack Urdu language support. The challenge was to build a full-stack SaaS platform that allows non-technical business owners to design, configure, and deploy a live conversational AI voice agent directly to their website in under 5 minutes, without writing a single prompt.
Process
CypherLM was architected to be the "Shopify for AI voice agents" — opinionated, fast, and built for non-technical users. We designed a guided 9-step wizard that captures business intent and compiles it into a structured system prompt via GPT-4o. The platform seamlessly integrates with Vapi for real-time telephony, Deepgram/Google for multilingual speech recognition, and ElevenLabs for human-quality TTS. A critical engineering milestone was supporting bilingual conversations (English + Urdu mid-call switching) and providing an isolated, self-contained <script> widget that can be embedded securely on any third-party domain.
Dual build pipeline
Engineered a dual Vite configuration (SPA + widget) producing a standalone, side-effect-free bundle for third-party embedding without polluting the host page's global namespace.
Multilingual voice routing
Implemented intelligent language-detection branches, routing to Deepgram for monolingual calls or Google Gemini STT for bilingual calls to enable accurate mid-sentence switching between English and Urdu.
Serverless DB caching
Optimized MongoDB connections in Vercel serverless functions with module-level caching, dropping cold-start API latency from 1.4s down to <120ms for hot paths.
Domain security & deployment
Built a secure embed architecture where domain verification happens server-side on every widget initialization, preventing unauthorized cross-deployment of agents.
Project Features
9-Step Guided Wizard & AI Prompt Compiler
Collects business context and uses GPT-4o to generate a production-grade system prompt automatically.
Multilingual Mid-Call Switching
Uses Google Gemini STT and ElevenLabs Multilingual v2 to seamlessly switch between English and Urdu phonemes.
Self-Contained Embed Widget
A dual-build Vite pipeline produces an isolated 1.4MB IIFE widget that drops into any website with two lines of HTML.
Domain Whitelisting & Admin Controls
Comprehensive admin dashboard and server-side Origin header checks to lock widgets securely to allowed domains.
Tech Stack
Outcomes
Conclusion
CypherLM successfully bridges the gap between advanced voice AI and non-technical business owners in Pakistan. By abstracting the complexity of prompt engineering, model orchestration, and WebRTC telephony into a beautifully guided experience, it empowers any SMB to deploy a production-grade, bilingual AI receptionist in minutes.
.avif)
