All work
Project lead & main engineer
Kumush AI: Multi-Tenant AI Voice Platform (Voice · Chat · Workflow)
2-month engagement · In production
A multi-tenant AI front desk platform where one agent handles customers across phone, chat, and automation, completing orders, forms, and API actions mid-conversation. Over a two-month engagement I led and built it as project lead and main engineer: planning, architecture, and latency optimization across the real-time voice stack, the embeddable chat widget, and the workflow / automation layer.
Highlights
- Real-time multilingual voice (Uzbek / Russian / English) on LiveKit; perceived response latency cut via MCP tool caching + TTS chunking.
- Voice stack built from modular STT → LLM → TTS layers, not a single speech-to-speech API, multi-provider with automatic model fallback.
- Inbound / outbound telephony (LiveKit, SIP/PSTN) plus SMS.
- Reception layer: forms and workflow automation: AI screening, calls, email, human review, Google Sheets writeback.
- RAG knowledge base and shared MCP tools across both chat and voice.
- Continuous reliability layer: automatic model fallback, latency observability logged to Postgres, migration versioning (dbmate).
- Benchmarked real-time STT providers and tuned custom NER for local place names.
In production
- 7,000+ customers served via Kumush Chat on Wakil AI, Central Asia’s first sovereign legal AI platform, citation-backed to official Uzbek legal sources (~3,000 daily active users).
- 90%+ conversation completion rate on a live client screening flow.
- Powers the admissions pipeline and outbound pre-interview voice campaigns for Humblebee AI Academy.
- 3 languages live in production.
